Category: Uncategorized

  • Renewable energy

    Renewable energy (also called green energy) is energy made from renewable natural resources that are replenished on a human timescale. The most widely used renewable energy types are solar energy, wind power, bioenergy/biomass, and hydropower. Bioenergy and geothermal power are also significant in some countries. Renewable energy installations can be large or small and are suited for both urban and rural areas. Renewable energy is often deployed together with further electrification. This has several benefits: electricity can move heat and vehicles efficiently and is clean at the point of consumption.[1][2] Variable renewable energy sources are those that have a fluctuating nature, such as wind power and solar power. In contrast, controllable renewable energy sources include dammed hydroelectricity, bioenergy, or geothermal power.

    For an energy to be renewable, it must be an endless source[3]. Nuclear power is therefore not considered renewable, due to its reliance on non-renewable fuel.[4]

    Renewable energy systems have rapidly become more efficient and cheaper over the past 40 years.[5] A large majority of newly installed worldwide electricity capacity is now renewable.[6] Renewable energy sources, such as solar and wind power, have seen significant cost reductions over the past decade, making them more competitive with traditional fossil fuels.[7] In some geographic localities, photovoltaic solar or onshore wind is the cheapest new-build electricity.[8] From 2011 to 2021, renewable energy grew from 20% to 28% of the global electricity supply. Power from the sun and wind accounted for most of this increase, growing from a combined 2% to 10%. Use of fossil energy shrank from 68% to 62%.[9] In 2025, renewables accounted for 34% of global electricity generation, up from 32% in 2024, and are projected to reach over 43% by 2030.[10][11] Many countries already have renewables contributing more than 20% of their total energy supply, with some generating over half or even all their electricity from renewable sources.[12][13]

    The main motivation to use renewable energy instead of fossil fuels is to reduce greenhouse gas emissions, which cause climate change. In general, renewable energy sources pollute much less than fossil fuels.[14] Renewables also cause much less air pollution than fossil fuels, improving public health, and are less noisy.[14] The International Energy Agency estimates that to achieve net zero emissions by 2050, 90% of global electricity will need to be generated by renewables.[15] The current pace of renewable expansion remains far from this required rate globally,[16] including in major economies with high financial capacities such as the G7 and the EU.[17]

    The deployment of renewable energy still faces obstacles, especially fossil fuel subsidies,[18] lobbying by incumbent power providers,[19] and local opposition to the use of land for renewable installations.[20][21] Like all mining, the extraction of minerals required for many renewable energy technologies also results in environmental damage.[22]

    Some also consider nuclear power a renewable power source, although this is controversial, as nuclear energy requires mining uranium, a nonrenewable resource.

    Overview
    See also: Lists of renewable energy topics

    Renewable energy sources, especially solar photovoltaic and wind, are generating an increasing share of electricity.[23]

    Coal, oil, and natural gas remain the primary global energy sources even as renewables have begun rapidly increasing.[24][25]
    Definition

    Global share of renewable electricity compared to other power sources, 1985 – 2025.
    Renewable energy is usually understood as energy harnessed from continuously occurring natural phenomena. The International Energy Agency defines it as “energy derived from natural processes that are replenished at a faster rate than they are consumed”. Solar power, wind power, hydroelectricity, geothermal energy, and biomass are widely agreed to be the main types of renewable energy.[26] Renewable energy often displaces conventional fuels in four areas: electricity generation, hot water/space heating, transportation, and rural (off-grid) energy services.[27]

    Although almost all forms of renewable energy cause much fewer carbon emissions than fossil fuels, the term is not synonymous with low-carbon energy. Some non-renewable sources of energy, such as nuclear power,[contradictory]generate almost no emissions, while some renewable energy sources can be very carbon-intensive, such as the burning of biomass if it is not offset by planting new plants.[14] Renewable energy is also distinct from sustainable energy, a more abstract concept that seeks to group energy sources based on their overall permanent impact on future generations of humans. For example, biomass is often associated with unsustainable deforestation.[28]

    Role in addressing climate change
    As part of the global effort to limit climate change, most countries have committed to net zero greenhouse gas emissions.[29] In practice, this means phasing out fossil fuels and replacing them with low-emissions energy sources.[14] This much needed process, coined as “low-carbon substitutions”[30] in contrast to other transition processes including energy additions, needs to be accelerated multiple times in order to successfully mitigate climate change.[30] At the 2023 United Nations Climate Change Conference, around three-quarters of the world’s countries set a goal of tripling renewable energy capacity by 2030.[31] The European Union aimed in 2022 to generate 40% of its electricity from renewables by the same year,[32] and by 2025, 47% of EU electricity generation was by renewables.[33]

    Other benefits
    Main article: Climate change mitigation § Co-benefits
    Renewable energy is more evenly distributed around the world than fossil fuels, which are concentrated in a limited number of countries.[34] It also brings health benefits by reducing air pollution caused by the burning of fossil fuels. The potential worldwide savings in health care costs have been estimated at trillions of dollars annually.[35]

    Intermittency
    Main article: Variable renewable energy

    Energy from sunlight or other renewable energy is converted to potential energy for storage in devices such as electric batteries. The stored potential energy is later converted to electricity that is added to the power grid, even when the original energy source is not available.

    Estimated power demand over a week in May 2012 and May 2020, Germany, showing the variability in solar and wind power both day-to-day and month-to-month
    The two most important forms of renewable energy, solar and wind, are intermittent energy sources: they are not available constantly, resulting in lower capacity factors. In contrast, fossil fuel power plants, nuclear power plants and hydropower are usually able to produce precisely the amount of energy an electricity grid requires at a given time. Solar energy can only be captured during the day, and ideally in cloudless conditions. Wind power generation can vary significantly not only day-to-day, but even month-to-month.[36] This poses a challenge when transitioning away from fossil fuels: energy demand will often be higher or lower than what renewables can provide.[37]

    In the medium-term, this variability may require keeping some gas-fired power plants or other dispatchable generation on standby[38][39] until there is enough energy storage, demand response, grid improvement, or base load power from non-intermittent sources. In the long-term, energy storage is an important way of dealing with intermittency.[40] Using diversified renewable energy sources and smart grids can also help flatten supply and demand.[41]

    Sector coupling of the power generation sector with other sectors may increase flexibility: for example the transport sector can be coupled by charging electric vehicles and sending electricity from vehicle to grid.[42] Similarly the industry sector can be coupled by hydrogen produced by electrolysis,[43] and the buildings sector by thermal energy storage for space heating and cooling.[44]

    Building overcapacity for wind and solar generation can help ensure sufficient electricity production even during poor weather. In optimal weather, it may be necessary to curtail energy generation if it is not possible to use or store excess electricity.[45]

    Electrical energy storage
    Main articles: Energy storage and Grid energy storage
    Electrical energy storage is a collection of methods used to store electrical energy. Electrical energy is stored during times when production (especially from intermittent sources such as wind power, tidal power, solar power) exceeds consumption, and returned to the grid when production falls below consumption. Pumped-storage hydroelectricity accounts for more than 85% of all grid power storage.[46] Batteries are increasingly being deployed for storage[47] and grid ancillary services[48] and for domestic storage.[49] Green hydrogen is a more economical means of long-term renewable energy storage, in terms of capital expenditures compared to pumped hydroelectric or batteries.[50][51]

    Energy supply security
    Two main renewable energy sources – solar power and wind power – are usually deployed in distributed generation architecture, which offers specific benefits and comes with specific risks.[52] Notable risks are associated with centralisation of 90% of the supply chains in a single country (China) in the photovoltaic sector.[53] Mass-scale installation of photovoltaic power inverters with remote control, security vulnerabilities and backdoors results in cyberattacks that can disable generation from millions of physically decentralised panels, resulting in disappearance of hundreds of gigawatts of installed power from the grid in one moment.[54][55] Similar attacks have targeted wind power farms through vulnerabilities in their remote control and monitoring systems.[56] The European NIS2 directive partially responds to these challenges by extending the scope of cybersecurity regulations to the energy generation market.[57] Recent analyses indicate that global solar photovoltaic capacity surpassed 1 terawatt in 2024, providing about 6–7% of global electricity supply.[58] Renewable energy infrastructure is also increasingly vulnerable to extreme weather events linked to climate change, such as heat waves, wildfires, severe storms, and flooding. Solar farms can experience reduced output during prolonged heat or smoke conditions, while wind turbines may require shutdowns during high-wind events or face damage from icing or wave action. These climate-related stresses can threaten the continuity of electricity supply in regions with high shares of variable renewables. As a result, governments and grid operators are adopting climate-resilience standards, hardening infrastructure, and developing emergency-response protocols to maintain energy security under more frequent and severe weather conditions.[59]

    Mainstream technologies

    Share of electricity from renewables

    Renewable energy capacity has steadily grown, led by solar photovoltaic power.[60]
    Solar energy
    Main articles: Solar energy, Solar power, and Outline of solar energy
    Installed capacity and other key design parameters Value and year
    Global electricity power generation capacity 2,391.6 GW (2025)[61]
    Global electricity power generation capacity annual growth rate 25% (2014–2023)[62]
    Share of global electricity generation 5.5% (2023)[63]
    Levelized cost per megawatt hour Utility-scale photovoltaics: USD 38.343 (2019)[64]
    Primary technologies Photovoltaics, concentrated solar power, solar thermal collector
    Main applications Electricity, water heating, heating, ventilation, air conditioning (HVAC)

    A small, rooftop PV system in Bonn, Germany

    Komekurayama photovoltaic power station in Kofu, Japan
    Solar power produced around 1.3 terrawatt-hours (TWh) worldwide in 2022,[12] representing 4.6% of the world’s electricity. Almost all of this growth has happened since 2010.[65] Solar energy can be harnessed anywhere that receives sunlight; however, the amount of solar energy that can be harnessed for electricity generation is influenced by weather conditions, geographic location and time of day.[66]

    There are two mainstream ways of harnessing solar energy: solar thermal, which converts solar energy into heat; and photovoltaics (PV), which converts it into electricity.[14] PV is far more widespread, accounting for around two thirds of the global solar energy capacity as of 2022.[67] It is also growing at a much faster rate, with 170 GW newly installed capacity in 2021,[68] compared to 25 GW of solar thermal.[67]

    Passive solar refers to a range of construction strategies and technologies that aim to optimize the distribution of solar heat in a building. Examples include solar chimneys,[14] orienting a building to the sun, using construction materials that can store heat, and designing spaces that naturally circulate air.[69]

    From 2020 to 2022, solar technology investments almost doubled from US$162 billion to US$308 billion, driven by the sector’s increasing maturity and cost reductions, particularly in solar photovoltaic (PV), which accounted for 90% of total investments. China and the United States were the main recipients, collectively making up about half of all solar investments since 2013. Despite reductions in Japan and India due to policy changes and COVID-19, growth in China, the United States, and a significant increase from Vietnam’s feed-in tariff program offset these declines. Globally, the solar sector added 714 gigawatts (GW) of solar PV and concentrated solar power (CSP) capacity between 2013 and 2021, with a notable rise in large-scale solar heating installations in 2021, especially in China, Europe, Turkey, and Mexico.[70] In 2023, global solar power capacity grew by nearly 30%, driven by falling panel prices and expanded government incentives worldwide.[71]

    Photovoltaics
    Main articles: Growth of photovoltaics, Solar power by country, and List of photovoltaic power stations

    Swanson’s law–stating that solar module prices have dropped about 20% for each doubling of installed capacity—defines the “learning rate” of solar photovoltaics.[72][73]
    A photovoltaic system, consisting of solar cells assembled into panels, converts light into electrical direct current via the photoelectric effect.[74][75]

    PV systems range from small, residential and commercial rooftop or building integrated installations,[76][77][78] to large utility-scale photovoltaic power station.[79][80][81] A household’s solar panels can either be used for just that household or, if connected to an electrical grid, can be aggregated with millions of others.[82][83][84]

    The first utility-scale solar power plant was built in 1982 in Hesperia, California by ARCO.[85][86] The plant was not profitable and was sold eight years later.[87] However, over the following decades, PV cells became significantly more efficient and cheaper.[88] As a result, PV adoption has grown exponentially since 2010.[89] Global capacity increased from 230 GW at the end of 2015 to 890 GW in 2021.[90] PV grew fastest in China between 2016 and 2021, adding 560 GW, more than all advanced economies combined.[91] In 2025, four of the ten biggest solar power stations were in China, including the biggest, Talatan Solar Park.[92]

    Solar panels are recycled to reduce electronic waste and create a source for materials that would otherwise need to be mined,[93] but such business is still small and work is ongoing to improve and scale-up the process.[94][95][96]

    Solar thermal
    Main article: Solar thermal energy
    Unlike photovoltaic cells that convert sunlight directly into electricity, solar thermal systems convert it into heat. They use mirrors or lenses to concentrate sunlight onto a receiver, which in turn heats a water reservoir. The heated water can then be used in homes. The advantage of solar thermal is that the heated water can be stored until it is needed, eliminating the need for a separate energy storage system.[97] Solar thermal power can also be converted to electricity by using the steam generated from the heated water to drive a turbine connected to a generator. However, because generating electricity this way is much more expensive than photovoltaic power plants, there are very few in use today.[98]

    Floatovoltaics
    Main article: Floating solar
    Floatovoltiacs, or floating solar panels, are solar panels floating on bodies of water. There are both positive and negative points to this. Some positive points are increased efficiency and price decrease of water space compared to land space. A negative point is that making floating solar panels could be more expensive.

    Agrivoltaics
    Main article: Agrivoltaics
    Agrivoltaics is where there is simultaneous use of land for energy production and agriculture. There are again both positive and negative points. A positive viewpoint is there is a better use of land, which leads to lower land costs. A negative viewpoint is it the plants grown underneath would have to be plants that can grow well under shade, such as Polka Dot Plant, Pineapple Sage, and Begonia.[99] Agrivoltaics not only optimizes land use and reduces costs by enabling dual revenue streams from both energy production and agriculture, but it can also help moderate temperatures beneath the panels, potentially reducing water loss and improving microclimates for crop growth. However, careful design and crop selection are crucial, as the shading effect may limit the types of plants that can thrive, necessitating the use of shade-tolerant species and innovative management practices.[100]

    Wind power
    Main articles: Wind power and Wind power by country

    Wind energy generation by region over time[101]

    Burbo, NW-England

    Sunrise at the Fenton Wind Farm in Minnesota, United States
    Installed capacity and other key design parameters Value and year
    Global electricity power generation capacity 1017.2 GW (2023)[102]
    Global electricity power generation capacity annual growth rate 13% (2014–2023)[103]
    Share of global electricity generation 7.8% (2023)[63]
    Levelized cost per megawatt hour Land-based wind: USD 30.165 (2019)[104]
    Primary technology Wind turbine, windmill
    Main applications Electricity, pumping water (windpump)
    Humans have harnessed wind energy since at least 3500 BC. Until the 20th century, it was primarily used to power ships, windmills and water pumps. Today, the vast majority of wind power is used to generate electricity using wind turbines.[14] Modern utility-scale wind turbines range from around 600 kW to 9 MW of rated power. The power available from the wind is a function of the cube of the wind speed, so as wind speed increases, power output increases up to the maximum output for the particular turbine.[105] Areas where winds are stronger and more constant, such as offshore and high-altitude sites, are preferred locations for wind farms.

    Wind-generated electricity met nearly 4% of global electricity demand in 2015, with nearly 63 GW of new wind power capacity installed. Wind energy was the leading source of new capacity in Europe, the US and Canada, and the second largest in China. In Denmark, wind energy met more than 40% of its electricity demand while Ireland, Portugal and Spain each met nearly 20%.[106]

    Globally, the long-term technical potential of wind energy is believed to be five times total current global energy production, or 40 times current electricity demand, assuming all practical barriers needed were overcome. This would require wind turbines to be installed over large areas, particularly in areas of higher wind resources, such as offshore, and likely also industrial use of new types of VAWT turbines in addition to the horizontal axis units currently in use. As offshore wind speeds average ~90% greater than that of land, offshore resources can contribute substantially more energy than land-stationed turbines.[107]

    Investments in wind technologies reached US$161 billion in 2020, with onshore wind dominating at 80% of total investments from 2013 to 2022. Offshore wind investments nearly doubled to US$41 billion between 2019 and 2020, primarily due to policy incentives in China and expansion in Europe. Global wind capacity increased by 557 GW between 2013 and 2021, with capacity additions increasing by an average of 19% each year.[70]

    Hydropower
    Main articles: Hydroelectricity and Hydropower

    The Three Gorges Dam for hydropower on the Yangtze River in China

    Three Gorges Dam and Gezhouba Dam, China
    Installed capacity and other key design parameters Value and year
    Global electricity power generation capacity 1,267.9 GW (2023)[108]
    Global electricity power generation capacity annual growth rate 1.9% (2014–2023)[109]
    Share of global electricity generation 14.3% (2023)[63]
    Levelized cost per megawatt hour USD 65.581 (2019)[110]
    Primary technology Dam
    Main applications Electricity, pumped storage, mechanical power
    Since water is about 800 times denser than air, even a slow flowing stream of water, or moderate sea swell, can yield considerable amounts of energy. Water can generate electricity with a conversion efficiency of about 90%, which is the highest rate in renewable energy.[111] There are many forms of water energy:

    Historically, hydroelectric power came from constructing large hydroelectric dams and reservoirs, which are still popular in developing countries.[112] The largest of them are the Three Gorges Dam (2003) in China and the Itaipu Dam (1984) built by Brazil and Paraguay.
    Small hydro systems are hydroelectric power installations that typically produce up to 50 MW of power. They are often used on small rivers or as a low-impact development on larger rivers. China is the largest producer of hydroelectricity in the world and has more than 45,000 small hydro installations.[113]
    Run-of-the-river hydroelectricity plants derive energy from rivers without the creation of a large reservoir. The water is typically conveyed along the side of the river valley (using channels, pipes or tunnels) until it is high above the valley floor, whereupon it can be allowed to fall through a penstock to drive a turbine. A run-of-river plant may still produce a large amount of electricity, such as the Chief Joseph Dam on the Columbia River in the United States.[114] However many run-of-the-river hydro power plants are micro hydro or pico hydro plants.
    Much hydropower is flexible, thus complementing wind and solar, as it not intermittent.[115] In 2021, the world renewable hydropower capacity was 1,360 GW.[91] Only a third of the world’s estimated hydroelectric potential of 14,000 TWh/year has been developed.[116][117] New hydropower projects face opposition from local communities due to their large impact, including relocation of communities and flooding of wildlife habitats and farming land.[118] High cost and lead times from permission process, including environmental and risk assessments, with lack of environmental and social acceptance are therefore the primary challenges for new developments.[119] It is popular to repower old dams thereby increasing their efficiency and capacity as well as quicker responsiveness on the grid.[120] Where circumstances permit existing dams such as the Russell Dam built in 1985 may be updated with “pump back” facilities for pumped-storage which is useful for peak loads or to support intermittent wind and solar power. Because dispatchable power is more valuable than VRE[121][122] countries with large hydroelectric developments such as Canada and Norway are spending billions to expand their grids to trade with neighboring countries having limited hydro.[123]

    Bioenergy
    Main article: Bioenergy
    Further information: Biomass (energy) § Environmental impacts
    Installed capacity and other key design parameters Value and year
    Global electricity generation capacity 150.3 GW (2023)[124]
    Global electricity generation capacity annual growth rate 5.8% (2014–2023)[125]
    Share of global electricity generation 2.4% (2022)[63]
    Levelized cost per megawatt hour USD 118.908 (2019)[126]
    Primary technologies Biomass, biofuel
    Main applications Electricity, heating, cooking, transportation fuels
    Biomass is biological material derived from living, or recently living organisms. Most commonly, it refers to plants or plant-derived materials. As an energy source, biomass can either be used directly via combustion to produce heat, or converted to a more energy-dense biofuel like ethanol. Wood is the most significant biomass energy source as of 2012[127] and is usually sourced from a trees cleared for silvicultural reasons or fire prevention. Municipal wood waste – for instance, construction materials or sawdust – is also often burned for energy.[128] The biggest per-capita producers of wood-based bioenergy are heavily forested countries like Finland, Sweden, Estonia, Austria, and Denmark.[129]

    Bioenergy can be environmentally destructive if old-growth forests are cleared to make way for crop production. In particular, demand for palm oil to produce biodiesel has contributed to the deforestation of tropical rainforests in Brazil and Indonesia.[130] In addition, burning biomass still produces carbon emissions, although much less than fossil fuels (39 grams of CO2 per megajoule of energy, compared to 75 g/MJ for fossil fuels).[131]

    Some biomass sources are unsustainable at current rates of exploitation (as of 2017).[132]

    Biofuel
    Main article: Biofuel
    See also: Ethanol fuel, Sustainable biofuel, and Issues relating to biofuels

    A CHP power station using wood to supply 30,000 households in France
    Biofuels are primarily used in transportation, providing 3.5% of the world’s transport energy demand in 2022,[133] up from 2.7% in 2010.[134] Biojet is expected to be important for short-term reduction of carbon dioxide emissions from long-haul flights.[135]

    A bus fueled by biodiesel
    Aside from wood, the major sources of bioenergy are bioethanol and biodiesel.[14] Bioethanol is usually produced by fermenting the sugar components of crops like sugarcane and maize, while biodiesel is mostly made from oils extracted from plants, such as soybean oil and corn oil.[136] Most of the crops used to produce bioethanol and biodiesel are grown specifically for this purpose,[137] although used cooking oil accounted for 14% of the oil used to produce biodiesel as of 2015.[136] The biomass used to produce biofuels varies by region. Maize is the major feedstock in the United States, while sugarcane dominates in Brazil.[138] In the European Union, where biodiesel is more common than bioethanol, rapeseed oil and palm oil are the main feedstocks.[139] China, although it produces comparatively much less biofuel, uses mostly corn and wheat.[140] In many countries, biofuels are either subsidized or mandated to be included in fuel mixtures.[130]

    Sugarcane plantation to produce ethanol in Brazil
    There are many other sources of bioenergy that are more niche, or not yet viable at large scales. For instance, bioethanol could be produced from the cellulosic parts of crops, rather than only the seed as is common today.[141] Sweet sorghum may be a promising alternative source of bioethanol, due to its tolerance of a wide range of climates.[142] Cow dung can be converted into methane.[143] There is also a great deal of research involving algal fuel, which is attractive because algae is a non-food resource, grows around 20 times faster than most food crops, and can be grown almost anywhere.[144]

    Geothermal energy
    Main articles: Geothermal energy, Geothermal power, Renewable thermal energy, and Geothermal energy in the United States

    Steam rising from the Nesjavellir Geothermal Power Station in Iceland

    Geothermal plant at The Geysers, California, US

    Krafla, a geothermal power station in Iceland
    Installed capacity and other key design parameters Value and year
    Global electricity power generation capacity 14.9 GW (2023)[145]
    Global electricity power generation capacity annual growth rate 3.4% (2014–2023)[146]
    Share of global electricity generation <1% (2018)[147]
    Levelized cost per megawatt hour USD 58.257 (2019)[148]
    Primary technologies Dry steam, flash steam, and binary cycle power stations
    Main applications Electricity, heating
    Geothermal energy is thermal energy (heat) extracted from the Earth’s crust. It originates from several different sources, of which the most significant is slow radioactive decay of minerals contained in the Earth’s interior,[14] as well as some leftover heat from the formation of the Earth.[149] Some of the heat is generated near the Earth’s surface in the crust, but some also flows from deep within the Earth from the mantle and core.[149] Geothermal energy extraction is viable mostly in countries located on tectonic plate edges, where the Earth’s hot mantle is more exposed.[150] As of 2023, the United States has by far the most geothermal capacity (2.7 GW,[151] or less than 0.2% of the country’s total energy capacity[152]), followed by Indonesia and the Philippines. Global capacity in 2022 was 15 GW.[151]

    Geothermal energy can be either used directly to heat homes, as is common in Iceland where almost all of its energy is renewable, or to generate electricity. Iceland is a global leader in renewable energy, relying almost entirely on its abundant geothermal and hydroelectric resources derived from volcanic activity and glaciers.[153] At smaller scales, geothermal power can be generated with geothermal heat pumps, which can extract heat from ground temperatures of under 30 °C (86 °F), allowing them to be used at relatively shallow depths of a few meters.[150] Electricity generation requires large plants and ground temperatures of at least 150 °C (302 °F). In some countries, electricity produced from geothermal energy accounts for a large portion of the total, such as Kenya (43%) and Indonesia (5%).[154]

    Technical advances may eventually make geothermal power more widely available. For example, enhanced geothermal systems involve drilling around 10 kilometres (6.2 mi) into the Earth, breaking apart hot rocks and extracting the heat using water. In theory, this type of geothermal energy extraction could be done anywhere on Earth.[150]

    Emerging technologies
    There are also other renewable energy technologies that are still under development, including enhanced geothermal systems, concentrated solar power, cellulosic ethanol, piezoelectricity, and marine energy.[155][156] These technologies are not yet widely demonstrated or have limited commercialization. Some may have potential comparable to other renewable energy technologies, but still depend on further breakthroughs from research, development and engineering.[156]

    Enhanced geothermal systems
    Main article: Enhanced geothermal systems
    Enhanced geothermal systems (EGS) are a new type of geothermal power which does not require natural hot water reservoirs or steam to generate power. Most of the underground heat within drilling reach is trapped in solid rocks, not in water.[157] EGS technologies use hydraulic fracturing to break apart these rocks and release the heat they contain, which is then harvested by pumping water into the ground. The process is sometimes known as “hot dry rock” (HDR).[158] Unlike conventional geothermal energy extraction, EGS may be feasible anywhere in the world, depending on the cost of drilling.[159] EGS projects have so far primarily been limited to demonstration plants, as the technology is capital-intensive due to the high cost of drilling.[160]

    Piezoelectricity
    Piezoelectricity is the conversion of existing mechanical stress or vibration (classical mechanics) into an electrical charge without consuming or depleting a fuel source.[161][162] Piezotronics enables the interaction of piezoelectric and semiconducting behaviors to modulate energy barriers at contact surface, thereby controlling charge carrier transport.[163] Since the introduction of nanogenerators, the efficiency of microscale energy harvesting has improved. For instance, nanogenerators typically consist of piezoelectric nanowires; as these wires bend or compress, the applied mechanical stress causes the ions within the material’s crystal lattice to shift their positions. This shift disrupts the nanowire’s charge symmetry which causes an instantaneous charge polarization (separation of positive and negative charges) across the nanowire’s ends. Once polarized, electrons are freed from the attached electrode which generates usable alternating current (AC) electricity that can energize low-power sensors.[164][165] Piezoelectric microelectromechanical systems (piezoMEMS), such as actuators for artificial organs and pacemakers or micropumps for drug delivery and reagent transfers, are vital for medical purposes and energy harvesting.[166] Furthermore, specialized components like piezoelectric resonators and quartz crystal oscillators are used to regulate electrical circuit frequencies.[167]

    Marine energy
    Main article: Marine energy

    Aerial view of Sihwa Tidal Power Station in South Korea
    Marine energy (also sometimes referred to as ocean energy) is the energy carried by ocean waves, tides, salinity, and ocean temperature differences. Technologies to harness the energy of moving water include wave power, marine current power, and tidal power. Reverse electrodialysis (RED) is a technology for generating electricity by mixing fresh water and salty sea water in large power cells.[168] Most marine energy harvesting technologies are still at low technology readiness levels and not used at large scales. Tidal energy is generally considered the most mature, but has not seen wide deployment.[169] The world’s largest tidal power station is on Sihwa Lake, South Korea,[170] which produces around 550 gigawatt-hours of electricity per year.[171]

    Earth infrared thermal radiation
    Earth emits roughly 1017 W of infrared thermal radiation that flows toward the cold outer space. Solar energy hits the surface and atmosphere of the earth and produces heat. Using various theorized devices like emissive energy harvester (EEH) or thermoradiative diode, this energy flow can be converted into electricity. In theory, this technology can be used during nighttime.[172][173]

    Others
    Algae fuels
    Main article: Algae fuels
    Producing liquid fuels from oil-rich (fat-rich) varieties of algae is an ongoing research topic. Various microalgae grown in open or closed systems are being tried including some systems that can be set up in brownfield and desert lands.[174]

    Space-based solar power
    Main article: Space-based solar power
    There have been numerous proposals for space-based solar power, in which very large satellites with photovoltaic panels would be equipped with microwave transmitters to beam power back to terrestrial receivers. A 2024 study by the NASA Office of Science and Technology Policy examined the concept and concluded that with current and near-future technologies it would be economically uncompetitive.[175]

    Water vapor
    Collection of static electricity charges from water droplets on metal surfaces is an experimental technology that would be especially useful in low-income countries with relative air humidity over 60%.[176]

    Nuclear energy
    Breeder reactors could, in principle, depending on the fuel cycle employed, extract almost all of the energy contained in uranium or thorium, decreasing fuel requirements by a factor of 100 compared to widely used once-through light water reactors, which extract less than 1% of the energy in the actinide metal (uranium or thorium) mined from the earth.[177] The high fuel-efficiency of breeder reactors could greatly reduce concerns about fuel supply, energy used in mining, and storage of radioactive waste. With seawater uranium extraction (currently too expensive to be economical), there is enough fuel for breeder reactors to satisfy the world’s energy needs for 5 billion years at 1983’s total energy consumption rate, thus making nuclear energy effectively a renewable energy.[178][179] In addition to seawater the average crustal granite rocks contain significant quantities of uranium and thorium with which breeder reactors can supply abundant energy for the remaining lifespan of the sun on the main sequence of stellar evolution.[180]

    Artificial photosynthesis
    Main article: Artificial photosynthesis
    Artificial photosynthesis uses techniques including nanotechnology to store solar electromagnetic energy in chemical bonds by splitting water to produce hydrogen and then using carbon dioxide to make methanol.[181] Researchers in this field strived to design molecular mimics of photosynthesis that use a wider region of the solar spectrum, employ catalytic systems made from abundant, inexpensive materials that are robust, readily repaired, non-toxic, stable in a variety of environmental conditions and perform more efficiently allowing a greater proportion of photon energy to end up in the storage compounds, i.e., carbohydrates (rather than building and sustaining living cells).[182] However, prominent research faces hurdles, Sun Catalytix a MIT spin-off stopped scaling up their prototype fuel-cell in 2012 because it offers few savings over other ways to make hydrogen from sunlight.[183]

    Recent research emphasizes that while artificial photosynthesis shows promise in splitting water to generate hydrogen, its broader significance lies in the ability to produce dense, carbon-based solar fuels suitable for transport applications, such as aviation and long-haul shipping. These fuels, if derived from carbon dioxide and water using sunlight, could close the carbon loop and reduce reliance on fossil-based hydrocarbons. However, realizing this potential requires overcoming major technical hurdles, including the development of efficient, durable catalysts for water oxidation and CO2 reduction, and careful attention to land use and public perception.[184]

    Technical potential
    Global energy consumption in 2019 was approximately 65 petawatt-hours (PWh).[185] The technical potentials for utility-scale solar photovoltaic, concentrated solar power, onshore wind, and offshore wind each exceed 100 PWh/year, and thus each of them is capable of meeting the total demand in theory.[186]

    Solar PV has a technical potential of around 5,800 PWh per year; almost 100 as much as is used annually.[187] The technical potential of wind energy (onshore + offshore) is nearly 900 PWh per year.[188] The theoretical potential of hydropower is about 52 PWh per year.[189] Conventional geothermal could be above 10 PWh/year,[186] enhanced geothermal systems could be about 4,000 PWh.[190]

    The literature assessing the global economic potential of renewables shows that the economic potential is higher than current and near-future electricity demand.[186] More specifically: Around 60% of the world’s solar resource and 15% of its wind resource is already economically competitive compared with local fossil fuel generation.[191]

    Market and industry trends
    Main article: Renewable energy commercialization

    Production of technology such as renewable energy sources starts a positive feedback loop to form what has been called a virtuous cycle—the opposite of a vicious cycle. For example, as more and more solar modules are deployed, prices fall because of the economies of scale, allowing the technology to become cost-competitive in new applications that in turn increase demand for more deployment.[192]

    Though the share of electricity generation from renewables has increased since 2010, there is substantial variance among major world regions.[193]

    In 2025, growth in solar, wind and other low-carbon electric power exceeded the overall growth in demand for electricity, reducing reliance on fossil fuels and helping to curb greenhouse gas emissions.[194]
    Most new renewables are solar, followed by wind then hydro then bioenergy.[195] Investment in renewables, especially solar, tends to be more effective in creating jobs than coal, gas or oil.[196][197] Worldwide, renewables employ about 12 million people as of 2020, with solar PV being the technology employing the most at almost 4 million.[198] However, as of February 2024, the world’s supply of workforce for solar energy is lagging greatly behind demand as universities worldwide still produce more workforce for fossil fuels than for renewable energy industries.[199]

    In 2021, China accounted for almost half of the global increase in renewable electricity.[200] There were 3,146 gigawatts installed in 135 countries, while 156 countries have laws regulating the renewable energy sector.[9][201]

    The International Renewable Energy Agency reported that during 2025 renewables accounted for 85.6% of new electricity generation capacity globally, with solar photovoltaics providing nearly three-quarters of the increase. At the end of 2025, renewables constituted 49.4% of global installed electricity generating capacity.[202][203]

    Globally in 2020 there are over 10 million jobs associated with the renewable energy industries, with solar photovoltaics being the largest renewable employer.[204] The clean energy sectors added about 4.7 million jobs globally between 2019 and 2022, totaling 35 million jobs by 2022.[205]: 5 

    Usage by sector or application
    Some studies say that a global transition to 100% renewable energy across all sectors – power, heat, transport and industry – is feasible and economically viable.[206][207][208]

    One of the efforts to decarbonize transportation is the increased use of electric vehicles (EVs).[209] Despite that and the use of biofuels, such as biojet, less than 4% of transport energy is from renewables.[210] Occasionally hydrogen fuel cells are used for heavy transport.[211] Meanwhile, in the future electrofuels may also play a greater role in decarbonizing hard-to-abate sectors like aviation and maritime shipping.[212]

    Solar water heating makes an important contribution to renewable heat in many countries, most notably in China, which now has 70% of the global total (180 GWth). Most of these systems are installed on multi-family apartment buildings[213] and meet a portion of the hot water needs of an estimated 50–60 million households in China. Worldwide, total installed solar water heating systems meet a portion of the water heating needs of over 70 million households.

    Heat pumps provide both heating and cooling, and also flatten the electric demand curve and are thus an increasing priority.[214] Renewable thermal energy is also growing rapidly.[215] About 10% of heating and cooling energy is from renewables.[216]

    Cost comparison
    The International Renewable Energy Agency (IRENA) stated that ~86% (187 GW) of renewable capacity added in 2022 had lower costs than electricity generated from fossil fuels.[217] IRENA also stated that capacity added since 2000 reduced electricity bills in 2022 by at least $520 billion, and that in non-OECD countries, the lifetime savings of 2022 capacity additions will reduce costs by up to $580 billion.[217] A 2025 evaluation consisting of 40 studies and over 150 scenarios found that cost projections for renewable energy technologies were primarily higher than observed market costs. These technologies particularly included utility-scale solar photovoltaics and lithium-ion battery storage. This suggests that actual costs have been significantly decreasing more than anticipated costs.[218]

    Installed[219]
    TWp Growth
    TW/yr[219] Production
    per installed
    capacity*[220] Energy
    TWh/yr*[220] Growth
    TWh/yr*[220] Levelized cost
    US¢/kWh[221] Av. auction prices
    US¢/kWh[222] Cost development
    2010–2019[221]
    Solar PV 0.580 0.098 13% 549 123 6.8 3.9 −82%
    Solar CSP 0.006 0.0006 13% 6.3 0.5 18.2 7.5 −47%
    Wind Offshore 0.028 0.0045 33% 68 11.5 11.5 8.2 −30%
    Wind Onshore 0.594 0.05 25% 1194 118 5.3 4.3 −38%
    Hydro 1.310 0.013 38% 4267 90 4.7 +27%
    Bioenergy 0.12 0.006 51% 522 27 6.6 −13%
    Geothermal 0.014 0.00007 74% 13.9 0.7 7.3 +49%

    • = 2018. All other values for 2019.

    Growth of renewables
    Investment and sources

    Investment: Companies, governments and households have committed increasing amounts to decarbonization, including renewable energy (solar, wind), electric vehicles and associated charging infrastructure, energy storage, energy-efficient heating systems, carbon capture and storage, and hydrogen.[223][224][225][226][227]

    By 2025, investment in the energy transition had grown to about twice that for fossil fuels (oil, natural gas and coal).[228]

    The countries most reliant on fossil fuels for electricity vary widely on how great a portion of that electricity is generated from renewables, leaving wide variation in renewables’ growth potential.[229]
    Costs

    Levelized cost: With increasingly widespread implementation of renewable energy sources, costs have declined, most notably for energy generated by solar panels.[230][231]
    Levelized cost of energy (LCOE) is a measure of the average net present cost of electricity generation for a generating plant over its lifetime.

    Costs of renewable energy, especially solar photovoltaic (solar panels), have declined significantly,[232] with 62% of total renewable power generation added in 2020 having lower costs than the cheapest new fossil fuel option.[233]

    “Learning curves”: Trend of costs and deployment over time, with steeper lines showing greater cost reductions as deployment progresses.[234] With increased deployment, renewables benefit from learning curves and economies of scale.[235]
    The results of a recent review of the literature concluded that as greenhouse gas (GHG) emitters begin to be held liable for damages resulting from GHG emissions resulting in climate change, a high value for liability mitigation would provide powerful incentives for deployment of renewable energy technologies.[236]

    In the decade of 2010–2019, worldwide investment in renewable energy capacity excluding large hydropower amounted to US$2.7 trillion, of which the top countries China contributed US$818 billion, the United States contributed US$392.3 billion, Japan contributed US$210.9 billion, Germany contributed US$183.4 billion, and the United Kingdom contributed US$126.5 billion.[237] This was an increase of over three and possibly four times the equivalent amount invested in the decade of 2000–2009 (no data is available for 2000–2003).[237]

    As of 2022, an estimated 28% of the world’s electricity was generated by renewables. This is up from 19% in 1990.[238] By the end of 2024, global renewable power capacity reached 4,300 gigawatts (GW), with solar photovoltaics accounting for over 60% of annual additions.[239]

    UK-based think tank Ember reported that wind and solar combined generated more electricity than gas globally in a month for the first time in April 2026.[240]

    Future projections
    See also: Energy transition

    Solar and wind power are scaling up faster than previous sources of electricity. Declining costs, modular design, and improved battery storage help the transition to renewable energy.[241]

    In 2023, electricity generation from wind and solar sources was projected to exceed 30% by 2030.[242]
    A December 2022 report by the IEA forecasts that over 2022–2027, renewables are seen growing by almost 2,400 GW in its main forecast, equal to the entire installed power capacity of China in 2021. This is an 85% acceleration from the previous five years, and almost 30% higher than what the IEA forecast in its 2021 report, making its largest ever upward revision. Renewables are set to account for over 90% of global electricity capacity expansion over the forecast period.[91] To achieve net zero emissions by 2050, IEA believes that 90% of global electricity generation will need to be produced from renewable sources.[21]

    In June 2022, IEA Executive Director Fatih Birol said that countries should invest more in renewables to “ease the pressure on consumers from high fossil fuel prices, make our energy systems more secure, and get the world on track to reach our climate goals.”[243]

    REPowerEU, the EU plan to escape dependence on fossil Russian gas, is expected to call for much more green hydrogen.[244]

    After a transitional period,[245] renewable energy production is expected to make up most of the world’s energy production. In 2018, the risk management firm, DNV GL, forecasts that the world’s primary energy mix will be split equally between fossil and non-fossil sources by 2050.[246]

    Middle eastern nations are also planning on reducing their reliance fossil fuel. Many planned green projects will contribute in 26% of energy supply for the region by 2050 achieving emission reductions equal to 1.1 Gt CO2/year.[247]

    Massive Renewable Energy Projects in the Middle East:[247]

    Mohammed bin Rashid Al Maktoum Solar Park in Dubai, UAE
    Shuaibah Two (2) Solar Facility in Mecca Province, Saudi Arabia
    NEOM Green Hydrogen Project in NEOM, Saudi Arabia
    Gulf of Suez Wind Power Project in Suez, Egypt
    Al-Ajban Solar Park in Abu Dhabi, UAE
    Demand
    In July 2014, the WWF and the World Resources Institute convened a discussion among a number of major US companies who had declared their intention to increase their use of renewable energy. These discussions identified a number of “principles” which companies seeking greater access to renewable energy considered important market deliverables. These principles included choice (between suppliers and between products), cost competitiveness, longer term fixed price supplies, access to third-party financing vehicles, and collaboration.[248]

    UK statistics released in September 2020 noted that “the proportion of demand met from renewables varies from a low of 3.4 per cent (for transport, mainly from biofuels) to highs of over 20 per cent for ‘other final users’, which is largely the service and commercial sectors that consume relatively large quantities of electricity, and industry”.[249]

    In some locations, individual households can opt to purchase renewable energy through a consumer green energy program.

    Developing countries
    This section is an excerpt from Renewable energy in developing countries.edit
    Renewable energy in developing countries is an increasingly used alternative to fossil fuel energy, as these countries scale up their energy supplies and address energy poverty. Renewable energy technology was once seen as unaffordable for developing countries.[250] However, since 2015, investment in non-hydro renewable energy has been higher in developing countries than in developed countries, and comprised 54% of global renewable energy investment in 2019.[251] The International Energy Agency forecasts that renewable energy will provide the majority of energy supply growth through 2030 in Africa and Central and South America, and 42% of supply growth in China.[252]

    Most developing countries have abundant renewable energy resources, including solar energy, wind power, geothermal energy, and biomass, as well as the ability to manufacture the relatively labor-intensive systems that harness these. By developing such energy sources developing countries can reduce their dependence on oil and natural gas, creating energy portfolios that are less vulnerable to price rises. In many circumstances, these investments can be less expensive than fossil fuel energy systems.[253]

    In Kenya, the Olkaria V Geothermal Power Station is one of the largest in the world.[254] The Grand Ethiopia Renaissance Dam project incorporates wind turbines.[255] Once completed, Morocco’s Ouarzazate Solar Power Station is projected to provide power to over a million people.[256]

    Policy

    Deaths caused as a result of fossil fuel use (areas of rectangles in chart) greatly exceed those resulting from production of renewable energy (rectangles barely visible in chart).[257]
    Policies to support renewable energy have been vital in their expansion. Where Europe dominated in establishing energy policy in the early 2000s, most countries around the world now have some form of energy policy.[258]

    The International Renewable Energy Agency (IRENA) is an intergovernmental organization for promoting the adoption of renewable energy worldwide. It aims to provide concrete policy advice and facilitate capacity building and technology transfer. IRENA was formed in 2009, with 75 countries signing the charter of IRENA.[259] As of April 2019, IRENA has 160 member states.[260] The then United Nations Secretary-General Ban Ki-moon has said that renewable energy can lift the poorest nations to new levels of prosperity.[261]

    The 2015 Paris Agreement on climate change motivated many countries to develop or improve renewable energy policies.[262] In 2017, a total of 121 countries adopted some form of renewable energy policy.[258] National targets that year existed in 176 countries.[262] In addition, there is also a wide range of policies at the state/provincial, and local levels.[134] Some public utilities help plan or install residential energy upgrades.

    Many national, state and local governments have created green banks. A green bank is a quasi-public financial institution that uses public capital to leverage private investment in clean energy technologies.[263] Green banks use a variety of financial tools to bridge market gaps that hinder the deployment of clean energy.

    Global and national policies related to renewable energy can be divided based on sectors, such as agriculture, transport, buildings, industry:

    Climate neutrality (net zero emissions) by the year 2050 is the main goal of the European Green Deal.[264] For the European Union to reach their target of climate neutrality, one goal is to decarbonise its energy system by aiming to achieve “net-zero greenhouse gas emissions by 2050.”[265]

    Finance

    Electrified transport and renewable energy are key areas of investment for climate change mitigation via an energy transition.[224][225][226]

    China’s count of multi-country patent filings has surged since the mid-2010s, leading in batteries and solar, though Europe still dominates in wind energy and smart grids.[266] China also leads in highly cited publications in global peer-reviewed journals.[266]
    The International Renewable Energy Agency’s (IRENA) 2023 report on renewable energy finance highlights steady investment growth since 2018: USD 348 billion in 2020 (a 5.6% increase from 2019), US$430 billion in 2021 (24% up from 2020), and US$499 billion in 2022 (16% higher). This trend is driven by increasing recognition of renewable energy’s role in mitigating climate change and enhancing energy security, along with investor interest in alternatives to fossil fuels. Policies such as feed-in tariffs in China and Vietnam have significantly increased renewable adoption. Furthermore, from 2013 to 2022, installation costs for solar photovoltaic (PV), onshore wind, and offshore wind dropped by 69%, 33%, and 45%, respectively, making renewables more cost-effective.[267][70]

    Between 2013 and 2022, the renewable energy sector underwent a significant realignment of investment priorities. Investment in solar and wind energy technologies markedly increased. In contrast, other renewable technologies such as hydropower (including pumped storage hydropower), biomass, biofuels, geothermal, and marine energy experienced a substantial decrease in financial investment. Notably, from 2017 to 2022, investment in these alternative renewable technologies declined by 45%, falling from US$35 billion to US$17 billion.[70]

    In 2023, the renewable energy sector experienced a significant surge in investments, particularly in solar and wind technologies, totaling approximately US$200 billion—a 75% increase from the previous year. The increased investments in 2023 contributed between 1% and 4% to the GDP in key regions including the United States, China, the European Union, and India.[268]

    The energy sector receives investments of approximately US$3 trillion each year, with US$1.9 trillion directed towards clean energy technologies and infrastructure. To meet the targets set in the Net Zero Emissions (NZE) Scenario by 2035, this investment must increase to US$5.3 trillion per year.[269]: 15 

    Debates
    Main articles: Renewable energy debate, Green job, and Intermittent power source
    Further information: Climate change mitigation § Overviews, strategies and comparisons of measures
    Nuclear power proposed as renewable energy

    The Leibstadt Nuclear Power Plant in Switzerland
    This section is an excerpt from Nuclear power as sustainable energy.edit
    Whether nuclear power should be considered a form of sustainable energy or renewable energy is an ongoing subject of debate. Legal definitions of renewable energy usually exclude nuclear energy technologies, with the notable exception of the U.S. state of Utah.[270] Dictionary-sourced definitions of renewable energy technologies often omit or explicitly exclude mention of nuclear energy sources, with an exception made for the natural nuclear decay heat generated within the Earth.[271][272]

    The most common fuel used in conventional nuclear fission power stations, uranium-235 is “non-renewable” according to the United States’ Energy Information Administration, the organization, however, is silent on the recycled MOX fuel.[272] The National Renewable Energy Laboratory does not mention nuclear power in its “energy basics” definition.[273]

    In 1987, the Brundtland Commission (WCED) classified fission reactors that produce more fissile nuclear fuel than they consume (breeder reactors, and if developed, fusion power) among conventional renewable energy sources, such as solar power and hydropower.[274] The monitoring and storage of radioactive waste products is also required upon the use of other renewable energy sources, such as geothermal energy.[275]

    Geopolitics
    See also: Electrostate
    Further information: Russia in the European energy sector

    A concept of a super grid
    The geopolitical impact of the growing use of renewable energy is a subject of ongoing debate and research.[276] Many countries wealthy in oil, such as Qatar, Russia, Saudi Arabia and Norway, are able to exert diplomatic or geopolitical influence as a result of their oil. Most of these countries are expected to be among the geopolitical “losers” of the energy transition, although some, like Norway, are also significant producers and exporters of renewable energy. Fossil fuels and the infrastructure to extract them may, in the long term, become stranded assets.[277] It has been speculated that countries dependent on fossil fuel revenue may one day find it in their interests to quickly sell off their remaining fossil fuels.[278]

    Conversely, nations abundant in renewable resources, and the minerals required for renewables technology, are expected to gain influence.[279][280] In particular, China has become the world’s dominant manufacturer of the technology needed to produce or store renewable energy, especially solar panels, wind turbines, and lithium-ion batteries.[281] Nations rich in solar and wind energy could become major energy exporters.[282] Some may produce and export green hydrogen,[283][282] although electricity is projected to be the dominant energy carrier in 2050, accounting for almost 50% of total energy consumption (up from 22% in 2015).[284] Countries with large uninhabited areas such as Australia, China, and many African and Middle Eastern countries have a potential for huge installations of renewable energy. The production of renewable energy technologies requires rare-earth elements with new supply chains.[285]

    Countries in Africa with already weak governments that rely on fossil fuel revenue may face even higher political instability or popular unrest. Analysts consider Nigeria, Angola, Chad, Gabon, and Sudan, all countries with a history of military coups, to be at risk of instability due to dwindling oil income.[286]

    A study found that transition from fossil fuels to renewable energy systems reduces risks from mining, trade and political dependence because renewable energy systems don’t need fuel – they depend on trade only for the acquisition of materials and components during construction.[287]

    In October 2021, European Commissioner for Climate Action Frans Timmermans suggested “the best answer” to the 2021 global energy crisis is “to reduce our reliance on fossil fuels.”[288] He said those blaming the European Green Deal were doing so “for perhaps ideological reasons or sometimes economic reasons in protecting their vested interests.”[288] Some critics blamed the European Union Emissions Trading System (EU ETS) and closure of nuclear plants for contributing to the energy crisis.[289][290][291] European Commission President Ursula von der Leyen said that Europe is “too reliant” on natural gas and too dependent on natural gas imports. According to Von der Leyen, “The answer has to do with diversifying our suppliers … and, crucially, with speeding up the transition to clean energy.”[292]

    Energy transition can also improve a country’s energy security and energy independence, mitigating impacts from fossil fuel markets or geopolitical pressure from oil producers.[293][294] The proliferation of renewable energy could also weaken petrostates’ trading power and their global influence. As the world becomes more electrified, renewable technologies will be embedded into the national economies of various markets, especially developing countries, removing the leverage petrostates have on the global energy structure.[295][296][297]

    Metal and mineral extraction
    See also: Environmental footprint of electric cars and Rare-earth element § Environmental considerations
    The transition to renewable energy requires increased extraction of certain metals and minerals. Like all mining, this impacts the environment[298] and can lead to environmental conflict.[299] For example, lithium mining uses around 65% of the water in the Salar de Atacama desert forcing farmers and llama herders to abandon their ancestral settlements and creating environment degradation,[300] in several African countries, the green energy transition has created a mining boom, causing deforestation, and threatening already endangered species.[301] Wind power requires large amounts of copper and zinc, as well as smaller amounts of the rarer metal neodymium. Solar power is less resource-intensive, but still requires significant amounts of aluminum. The expansion of electrical grids requires both copper and aluminum. Batteries, which are critical to enable storage of renewable energy, use large quantities of copper, nickel, aluminum and graphite. Demand for lithium is expected to grow 42-fold from 2020 to 2040. Demand for nickel, cobalt and graphite is expected to grow by a factor of about 20–25.[302] For each of the most relevant minerals and metals, its mining is dominated by a single country: copper in Chile, nickel in Indonesia, rare earths in China, cobalt in the Democratic Republic of the Congo (DRC), and lithium in Australia. China dominates processing of all of these.[302]

    Recycling these metals after the devices they are embedded in are spent is essential to create a circular economy and ensure renewable energy is sustainable. By 2040, recycled copper, lithium, cobalt, and nickel from spent batteries could reduce combined primary supply requirements for these minerals by around 10%.[302]

    A controversial approach is deep sea mining. Minerals can be collected from new sources like polymetallic nodules lying on the seabed.[303] This would damage local biodiversity,[304] but proponents point out that biomass on resource-rich seabeds is much scarcer than in the mining regions on land, which are often found in vulnerable habitats like rainforests.[305]

    Due to co-occurrence of rare-earth and radioactive elements (thorium, uranium and radium), rare-earth mining results in production of low-level radioactive waste.[306]

    Conservation areas
    Installations used to produce wind, solar and hydropower are an increasing threat to key conservation areas, with facilities built in areas set aside for nature conservation and other environmentally sensitive areas. They are often much larger than fossil fuel power plants, needing areas of land up to 10 times greater than coal or gas to produce equivalent energy amounts.[307] More than 2000 renewable energy facilities are built, and more are under construction, in areas of environmental importance and threaten the habitats of plant and animal species across the globe. The authors’ team emphasized that their work should not be interpreted as anti-renewables because renewable energy is crucial for reducing carbon emissions. The key is ensuring that renewable energy facilities are built in places where they do not damage biodiversity.[308]

    In 2020 scientists published a world map of areas that contain renewable energy materials as well as estimations of their overlaps with “Key Biodiversity Areas”, “Remaining Wilderness” and “Protected Areas”. The authors assessed that careful strategic planning is needed.[309][310][311]

    Impact of climate change on renewable energy production
    Climate change is making weather patterns less predictable. This can seriously hamper the use of renewable energy. For example, in the year 2023, in Sudan and Namibia, hydropower production dropped by more than half due to drastic reduction in rainfall, in China, India and some regions in Africa unusual weather phenomena reduced the amount of produced wind energy, heatwaves and clouds reduce the effectiveness of solar panels, melting glaciers create problems to hydropower. Nuclear energy is also affected as drought creates water shortage, so nuclear power plants sometimes do not have enough water for cooling.[312]

    Society and culture
    Public support

    Most survey respondents in Europe support renewable energy to fight climate change.[313]

    Survey respondents in the US considered investment in renewable energy to be a lower priority than those in the European Union and China.[314]

    Acceptance of wind and solar facilities in one’s community is stronger among U.S. Democrats (blue), while acceptance of nuclear power plants is stronger among U.S. Republicans (red).[315]
    Solar power plants may compete with arable land,[316][317] while on-shore wind farms often face opposition due to aesthetic concerns and noise.[318][319] Such opponents are often described as NIMBYs (“not in my back yard”).[320] Some environmentalists are concerned about fatal collisions of birds and bats with wind turbines.[321] Although protests against new wind farms occasionally occur around the world, regional and national surveys generally find broad support for both solar and wind power.[322][323][324] Opponents of renewable energy sometimes receive support from the fossil fuel industry and spread misinformation to undermine public support for renewable energy.[325][326][327]

    Community-owned wind energy is sometimes proposed as a way to increase local support for wind farms.[328] A 2011 UK Government document stated that “projects are generally more likely to succeed if they have broad public support and the consent of local communities. This means giving communities both a say and a stake.”[329] In the 2000s and early 2010s, many renewable projects in Germany, Sweden and Denmark were owned by local communities, particularly through cooperative structures.[330][331] In the years since, more installations in Germany have been undertaken by large companies,[328] but community ownership remains strong in Denmark.[332]

    History
    See also: History of climate change policy and politics
    Prior to the development of coal in the mid 19th century, nearly all energy used was renewable. The oldest known use of renewable energy, in the form of traditional biomass to fuel fires, dates from more than a million years ago. The use of biomass for fire did not become commonplace until many hundreds of thousands of years later.[333] The second oldest usage of renewable energy was probably harnessing the wind in order to drive ships over water. This practice can be traced back some 7000 years, to ships in the Persian Gulf and on the Nile.[334] Geothermal energy from hot springs has been used for bathing since Paleolithic times and space heating since ancient Roman times.[335] Moving into the time of recorded history, the primary sources of traditional renewable energy were human labor, animal power, water power, wind, in grain crushing windmills, and firewood, a traditional biomass.

    In 1885, Werner Siemens, commenting on the discovery of the photovoltaic effect in the solid state, wrote:

    In conclusion, I would say that however great the scientific importance of this discovery may be, its practical value will be no less obvious when we reflect that the supply of solar energy is both without limit and without cost, and that it will continue to pour down upon us for countless ages after all the coal deposits of the earth have been exhausted and forgotten.[336]

    Max Weber mentioned the end of fossil fuel in the concluding paragraphs of his Die protestantische Ethik und der Geist des Kapitalismus (The Protestant Ethic and the Spirit of Capitalism), published in 1905.[337] Development of solar engines continued until the outbreak of World War I. The importance of solar energy was recognized in a 1911 Scientific American article: “in the far distant future, natural fuels having been exhausted [solar power] will remain as the only means of existence of the human race”.[338]

    The theory of peak oil was published in 1956.[339] In the 1970s environmentalists promoted the development of renewable energy both as a replacement for the eventual depletion of oil, as well as for an escape from dependence on oil, and the first electricity-generating wind turbines appeared. Solar had long been used for heating and cooling, but solar panels were too costly to build solar farms until 1980.[340]

    New government spending, regulation and policies helped the renewables industry weather the 2008 financial crisis and the Great Recession better than many other sectors.[341] In 2022, renewables accounted for 30% of global electricity generation, up from 21% in 1985.[342]

    Ancient Historical Examples
    Among the most notable historical uses of renewable energy (in the form of ancient and traditional methods), the following examples can be highlighted:

    Windmills in Europe and Asia (such as the windmills of the Netherlands and Nashtifan in Iran).[343] The earliest discovered verified designs of windmills date back to Iran, between 700 and 900 CE.[344][345][346]
    Water mills (Ancient China and Ancient Persia).[347]
    Archimedes’ burning lens.
    Traditional cooling and ventilation systems based on windcatchers and Solar updraft tower (or Solar chimney).
    Traditional architecture aware of natural heat transfer and natural energy transformation processes.
    Gravity-based fountains.
    Using animal biomass in ancient fuel bricks.
    Solar ovens and furnaces in ancient China, India, Egypt, and Persia.
    Solar energy applications for traditional agricultural processing (drying), engineering material properties (solar curing of pottery and ceramics), and ancient health practices (natural disinfection by solar radiation).
    Long-distance gravitational water flow control in ancient qanat technology for water transport and supply.
    Cargo and passenger transportation using sails on rivers, seas, and oceans.
    Cargo and passenger transportation based on understanding water currents in rivers, seas, and oceans.
    Using renewable vegetation (such as desert shrubs, agricultural waste, and pruned branches) for producing light and heat.
    Using renewable oils (vegetable or animal-based) for producing light and heat.
    Maximizing use of natural sunlight during the day and moonlight at night in building architecture for purposes such as lighting, decorative applications (e.g., reflective tilework, mirror work, and surface polishing on stone or brick), timekeeping (sundials, noon markers, prayer time indicators, seasonal change markers),

  • Climate change

    Present-day climate change includes both global warming—the ongoing increase in global average temperature—and its wider effects on Earth’s climate system. In a broader sense, climate change also includes previous long-term changes to Earth’s climate. The modern-day rise in global temperatures is driven by human activities, especially fossil fuel (coal, oil and natural gas) burning since the Industrial Revolution.[4][5] Fossil fuel use, deforestation, and some agricultural and industrial practices release greenhouse gases.[6] These gases absorb some of the heat that the Earth radiates after it warms from sunlight, warming the lower atmosphere. Earth’s atmosphere now has roughly 50% more carbon dioxide, the main gas driving global warming, than it did at the end of the pre-industrial era, reaching levels not seen for millions of years.[7]

    Climate change has an increasingly large impact on the environment. Heat waves and wildfires are becoming more common.[8] Amplified warming in the Arctic has contributed to thawing permafrost, retreat of glaciers and sea ice decline.[9] Higher temperatures are also causing more intense storms, droughts, and other weather extremes.[10] Rapid environmental change in mountains, coral reefs, and the Arctic is forcing many species to relocate or become extinct.[11] Even if efforts to minimize future warming are successful, some effects will continue for centuries. These include ocean heating, ocean acidification and sea level rise.[12]

    Climate change threatens people with increased flooding, extreme heat, increased food and water scarcity, more disease, and economic loss.[13] Human migration and conflict can also be a result.[14] The World Health Organization calls climate change one of the biggest threats to global health in the 21st century.[15] Societies and ecosystems will experience more severe risks without action to limit warming.[16] Adapting to climate change through efforts like flood control measures or drought-resistant crops partially reduces climate change risks, although some limits to adaptation have already been reached.[17] Poorer communities are responsible for a small share of global emissions, yet have the least ability to adapt and are most vulnerable to climate change.[18][19]

    Bobcat Fire in Monrovia, CA, September 10, 2020
    Bleached colony of Acropora coral
    A dry lakebed in California, which is experiencing its worst megadrought in 1,200 years.
    Examples of some effects of climate change: Wildfire intensified by heat and drought, bleaching of corals occurring more often due to marine heatwaves, and worsening droughts compromising water supplies.
    Many climate change impacts have been observed in the first decades of the 21st century, with 2024 the warmest on record at +1.60 °C (2.88 °F) since regular tracking began in 1850.[21][22] Additional warming will increase these impacts and can trigger tipping points, such as melting all of the Greenland ice sheet.[23] Under the 2015 Paris Agreement, nations collectively agreed to keep warming “well under 2 °C”. However, with pledges made under the Agreement, global warming would still reach about 2.8 °C (5.0 °F) by the end of the century.[24]

    There is widespread support for climate action worldwide,[25][26] and most countries aim to stop emitting carbon dioxide.[27] Fossil fuels can be phased out by ending their subsidies, conserving energy and switching to energy sources that do not produce significant carbon pollution. These energy sources include wind, solar, hydro, and nuclear power.[28] Cleanly generated electricity can replace fossil fuels for powering transportation, heating buildings, and running industrial processes.[29] Carbon can also be removed from the atmosphere, for instance by increasing forest cover and farming with methods that store carbon in soil.[30][31][32]

    Terminology
    Before the 1980s, it was unclear whether the warming effect of increased greenhouse gases was stronger than the cooling effect of airborne particulates in air pollution. Scientists used the term inadvertent climate modification to refer to human impacts on the climate at this time.[33] In the 1980s, the terms global warming and climate change became more common, often being used interchangeably.[34][35][36] Scientifically, global warming refers only to increased global average surface temperature, while climate change describes both global warming and its effects on Earth’s climate system, such as precipitation changes.[33]

    Climate change can also be used more broadly to include changes to the climate that have happened throughout Earth’s history as result of natural processes.[37] The term anthropogenic climate change is sometimes used to describe climate change resulting from human activities.[38]

    Global warming—used as early as 1975[39]—became the more popular term after NASA climate scientist James Hansen used it in his 1988 testimony in the U.S. Senate.[40] Since the 2000s, usage of climate change has increased.[41] Various scientists, politicians and media may use the terms climate crisis or climate emergency to talk about climate change, and may use the term global heating instead of global warming.[42][43]

    Global temperature rise
    Further information: Global surface temperature
    Temperatures prior to present-day global warming
    Main articles: Climate variability and change; Temperature record of the last 2,000 years; and Paleoclimatology

    Global surface temperature reconstruction over the past 2000 years using proxy data from tree rings, corals, and ice cores in blue.[44] Directly observed data is in red.[45]
    Over the last few million years the climate cycled through ice ages. One of the hotter periods was the Last Interglacial, around 125,000 years ago, where temperatures were between 0.5 °C and 1.5 °C warmer than before the start of global warming.[46] This period saw sea levels 5 to 10 metres higher than today. The most recent glacial maximum 20,000 years ago was some 5–7 °C colder. This period has sea levels that were over 125 metres (410 ft) lower than today.[47]

    Temperatures stabilized in the current interglacial period beginning 11,700 years ago.[48] This period also saw the start of agriculture.[49] Historical patterns of warming and cooling, like the Medieval Warm Period and the Little Ice Age, did not occur at the same time across different regions. Temperatures may have reached as high as those of the late 20th century in a limited set of regions.[50][51] Climate information for that period comes from climate proxies, such as trees and ice cores.[52][53]

    Warming since the Industrial Revolution

    In recent decades, new high temperature records have substantially outpaced new low temperature records on a growing portion of Earth’s surface.[54]

    There has been an increase in ocean heat content during recent decades as the oceans absorb over 90% of the heat from global warming.[55]
    Around 1850 thermometer records began to provide global coverage.[56] Between the 18th century and 1970 there was little net warming, as the warming impact of greenhouse gas emissions was offset by cooling from sulfur dioxide emissions. Sulfur dioxide causes acid rain, but it also produces sulfate aerosols in the atmosphere, which reflect sunlight and cause global dimming. After 1970, the increasing accumulation of greenhouse gases and controls on sulfur pollution led to a marked increase in temperature.[57][58]

    Duration: 30 seconds.0:30
    NASA animation portraying global surface temperature changes since 1880. The colour blue denotes cooler temperatures and red denotes warmer temperatures. As reference value the mean temperature from 1951 to 1980 is used.
    Ongoing changes in climate have had no precedent for several thousand years.[59] Multiple datasets all show worldwide increases in surface temperature, at a rate of around 0.2 °C per decade.[60] The 2016–2025 decade warmed to an average 1.26 °C [1.13–1.36 °C] compared to the pre-industrial baseline (1850–1900).[61] Not every single year was warmer than the last: internal climate variability processes can make any year 0.2 °C warmer or colder than the average.[62] From 1998 to 2013, negative phases of two such processes, Pacific decadal oscillation (PDO)[63] and Atlantic multidecadal oscillation (AMO)[64] caused a short slower period of warming called the “global warming hiatus”.[65] After the “hiatus”, the opposite occurred, with 2024 well above the recent average at more than +1.5 °C.[66] This is why the temperature change is defined in terms of a 20-year average, which reduces the noise of hot and cold years and decadal climate patterns, and detects the long-term signal.[67]: 5 [68]

    A wide range of other observations reinforce the evidence of warming.[69][70] The upper atmosphere is cooling, because greenhouse gases are trapping heat near the Earth’s surface, and so less heat is radiating into space.[71] Warming reduces average snow cover and forces the retreat of glaciers. At the same time, warming also causes greater evaporation from the oceans, leading to more atmospheric humidity, and more and heavier precipitation.[72][73] Plants are flowering earlier in spring, and thousands of animal species have been permanently moving to cooler areas.[74]

    Differences by region
    Different regions of the world warm at different rates. The pattern is independent of where greenhouse gases are emitted, because the gases persist long enough to diffuse across the planet. Since the pre-industrial period, the average surface temperature over land regions has increased almost twice as fast as the global average surface temperature.[75] This is because oceans lose more heat by evaporation and oceans can store a lot of heat.[76] The thermal energy in the global climate system has grown with only brief pauses since at least 1970, and over 90% of this extra energy has been stored in the ocean.[77][78] The rest has heated the atmosphere, melted ice, and warmed the continents.[79]

    The Northern Hemisphere and the North Pole have warmed much faster than the South Pole and Southern Hemisphere. The Northern Hemisphere not only has much more land, but also more seasonal snow cover and sea ice. As these surfaces flip from reflecting a lot of light to being dark after the ice has melted, they start absorbing more heat.[80] Local black carbon deposits on snow and ice also contribute to Arctic warming.[81] Arctic surface temperatures are increasing between three and four times faster than in the rest of the world.[82][83] Melting of ice sheets near the poles weakens both the Atlantic and the Antarctic limb of thermohaline circulation, which further changes the distribution of heat and precipitation around the globe.[84][85][86][87]

    Future global temperatures

    CMIP6 multi-model projections of global surface temperature changes for the year 2090 relative to the 1850–1900 average. The current trajectory for warming by the end of the century is roughly halfway between these two extremes.[24][88][89]
    The World Meteorological Organization estimates there is almost a 50% chance of the five-year average global temperature exceeding +1.5 °C between 2024 and 2028.[90] The IPCC expects the 20-year average to exceed +1.5 °C in the early 2030s.[91]

    The IPCC Sixth Assessment Report (2021) included projections that by 2100 global warming is very likely to reach 1.0–1.8 °C under a scenario with very low emissions of greenhouse gases, 2.1–3.5 °C under an intermediate emissions scenario, or 3.3–5.7 °C under a very high emissions scenario.[92] The warming will continue past 2100 in the intermediate and high emission scenarios,[93][94] with future projections of global surface temperatures by year 2300 being similar to millions of years ago.[95]

    The remaining carbon budget for staying beneath certain temperature increases is determined by modelling the carbon cycle and climate sensitivity to greenhouse gases.[96] According to UNEP, global warming can be kept below 2.0 °C with a 50% chance if emissions after 2023 do not exceed 900 gigatonnes of CO2. This carbon budget corresponds to around 16 years of current emissions.[97]

    Causes of recent global temperature rise
    Main article: Causes of climate change

    Physical drivers of global warming that has happened so far. Future global warming potential for long lived drivers like carbon dioxide emissions is not represented. Whiskers on each bar show the possible error range.
    The climate system experiences various cycles on its own which can last for years, decades or even centuries. For example, El Niño events cause short-term spikes in surface temperature while La Niña events cause short term cooling.[98] Their relative frequency can affect global temperature trends on a decadal timescale.[99] Other changes are caused by an imbalance of energy from external forcings.[100] Examples of these include changes in the concentrations of greenhouse gases, solar luminosity, volcanic eruptions, and variations in the Earth’s orbit around the Sun.[101]

    To determine the human contribution to climate change, unique “fingerprints” for all potential causes are developed and compared with both observed patterns and known internal climate variability.[102] For example, solar forcing—whose fingerprint involves warming the entire atmosphere—is ruled out because only the lower atmosphere has warmed.[103] Atmospheric aerosols produce a smaller, cooling effect. Other drivers, such as changes in albedo, are less impactful.[104]

    Greenhouse gases
    Main articles: Greenhouse gas, Greenhouse gas emissions, Greenhouse effect, and Carbon dioxide in Earth’s atmosphere

    CO2 concentrations over the last 800,000 years as measured from ice cores (blue/green) and directly (black)
    Greenhouse gases are transparent to sunlight, and thus allow it to pass through the atmosphere to heat the Earth’s surface. The Earth radiates it as heat, and greenhouse gases absorb a portion of it. This absorption slows the rate at which heat escapes into space, trapping heat near the Earth’s surface and warming it over time.[105]

    While water vapour (≈50%) and clouds (≈25%) are the biggest contributors to the greenhouse effect, they primarily change as a function of temperature and are therefore mostly considered to be feedbacks that change climate sensitivity. On the other hand, concentrations of gases such as CO2 (≈20%), tropospheric ozone,[106] CFCs and nitrous oxide are added or removed independently from temperature, and are therefore considered to be external forcings that change global temperatures.[107]

    Before the Industrial Revolution, naturally occurring amounts of greenhouse gases caused the air near the surface to be about 33 °C warmer than it would have been in their absence.[108][109] Human activity since the Industrial Revolution, mainly extracting and burning fossil fuels (coal, oil, and natural gas),[110] has increased the amount of greenhouse gases in the atmosphere. In 2022, the concentrations of CO2 and methane had increased by about 50% and 164%, respectively, since 1750.[111] These CO2 levels are higher than they have been at any time during the last 14 million years, i.e., during the middle Cenozoic.[112] Concentrations of methane are far higher than they were over the last 800,000 years.[113]

    The Global Carbon Project shows how additions to CO2 have been caused by different sources ramping up one after another.[114]
    Global human-caused greenhouse gas emissions in 2019 were equivalent to 59 billion tonnes of CO2. Of these emissions, 75% was CO2, 18% was methane, 4% was nitrous oxide, and 2% was fluorinated gases.[115] CO2 emissions primarily come from burning fossil fuels to provide energy for transport, manufacturing, heating, and electricity.[6] Additional CO2 emissions come from deforestation and industrial processes, which include the CO2 released by the chemical reactions for making cement, steel, aluminium, and fertilizer.[116][117][118][119] Methane emissions come from livestock, wetlands, anure, rice cultivation, landfills, wastewater, and coal mining, as well as oil and gas extraction.[120][121] Nitrous oxide emissions largely come from the microbial decomposition of fertilizer.[122][123]

    While methane only lasts in the atmosphere for an average of 12 years,[124] CO2 lasts much longer. The Earth’s surface absorbs CO2 as part of the carbon cycle. While plants on land and in the ocean absorb most excess emissions of CO2 every year, that CO2 is returned to the atmosphere when biological matter is digested, burns, or decays.[125] Land-surface carbon sink processes, such as carbon fixation in the soil and photosynthesis, remove about 29% of annual global CO2 emissions.[126] The ocean has absorbed 20 to 30% of emitted CO2 over the last two decades.[127] CO2 is only removed from the atmosphere for the long term when it is stored in the Earth’s crust, which is a process that can take millions of years to complete.[125]

    Land surface changes

    The rate of global tree cover loss has approximately doubled since 2001, to an annual loss approaching an area the size of Italy.[128]
    Around 30% of Earth’s land area is largely unusable for humans (glaciers, deserts, etc.), 26% is forests, 10% is shrubland and 34% is agricultural land.[129] Deforestation is the main land use change contributor to global warming,[130] as the destroyed trees release CO2, and are not replaced by new trees, removing that carbon sink.[131] Between 2001 and 2018, 27% of deforestation was from permanent clearing to enable agricultural expansion for crops and livestock. Another 24% has been lost to temporary clearing under the shifting cultivation agricultural systems. 26% was due to logging for wood and derived products, and wildfires have accounted for the remaining 23%.[132] Some forests have not been fully cleared, but were already degraded by these impacts. Restoring these forests also recovers their potential as a carbon sink.[133]

    Local vegetation cover impacts how much of the sunlight gets reflected back into space (albedo), and how much heat is lost by evaporation. For instance, the change from a dark forest to grassland makes the surface lighter, causing it to reflect more sunlight. Deforestation can also modify the release of chemical compounds that influence clouds, and can change the roughness of Earth’s surface in a way affecting wind speed.[134] In tropic and temperate areas the net effect is to produce significant warming, and forest restoration can make local temperatures cooler.[133] At latitudes closer to the poles, there is a cooling effect as forest is replaced by snow-covered (and more reflective) plains.[134] Globally, these increases in surface albedo have been the dominant direct influence on temperature from land use change. Thus, land use change to date is estimated to have a slight cooling effect.[135]

    Other factors
    Aerosols and clouds
    Air pollution, in the form of aerosols, affects the climate on a large scale.[136] Aerosols scatter and absorb solar radiation. From 1961 to 1990, a gradual reduction in the amount of sunlight reaching the Earth’s surface was observed. This phenomenon is popularly known as global dimming,[137] and is primarily attributed to sulfate aerosols produced by the combustion of fossil fuels with heavy sulfur concentrations like coal and bunker fuel.[58] Smaller contributions come from black carbon (from combustion of fossil fuels and biomass), and from dust.[138][139][140] Globally, aerosols have been declining since 1990 due to pollution controls, meaning that they no longer mask greenhouse gas warming as much.[141][58]

    Aerosols also have indirect effects on the Earth’s energy budget. Sulfate aerosols act as cloud condensation nuclei and lead to clouds that have more and smaller cloud droplets. These clouds reflect solar radiation more efficiently than clouds with fewer and larger droplets.[142] They also reduce the growth of raindrops, which makes clouds more reflective to incoming sunlight.[143] Indirect effects of aerosols are the largest uncertainty in radiative forcing.[144]

    While aerosols typically limit global warming by reflecting sunlight, black carbon in soot that falls on snow or ice can contribute to global warming. The resulting excess heat accelerates the melting of ice sheets and glaciers, which in turn contributes significantly to global sea-level rise.[145][146][147][81] Limiting new black carbon deposits in the Arctic could reduce global warming by 0.2 °C by 2050.[148] The effect of decreasing sulfur content of fuel oil for ships since 2020[149] is estimated to cause an additional 0.05 °C increase in global mean temperature by 2050.[150]

    Solar and volcanic activity
    Further information: Solar activity and climate

    The Fourth National Climate Assessment (“NCA4”, USGCRP, 2017) includes charts illustrating that neither solar nor volcanic activity can explain the observed warming.[151][152]
    As the Sun is the Earth’s primary energy source, changes in incoming sunlight directly affect the climate system.[144] Solar irradiance has been measured directly by satellites,[153] and indirect measurements are available from the early 1600s onwards.[144] Since 1880, there has been no upward trend in the amount of the Sun’s energy reaching the Earth, in contrast to the warming of the lower atmosphere (the troposphere).[154] The upper atmosphere (the stratosphere) would also be warming if the Sun was sending more energy to Earth, but instead, it has been cooling.[103] This is consistent with greenhouse gases preventing heat from leaving the Earth’s atmosphere.[155]

    Explosive volcanic eruptions can release gases, dust and ash that partially block sunlight and reduce temperatures, or they can send water vapour into the atmosphere, which adds to greenhouse gases and increases temperatures. These impacts on temperature only last for several years, because both water vapour and volcanic material have low persistence in the atmosphere.[156] Volcanic CO2 emissions are more persistent, but they are equivalent to less than 1% of current human-caused CO2 emissions.[157] Volcanic activity still represents the single largest natural impact (forcing) on temperature in the industrial era. Yet, like the other natural forcings, it has had negligible impacts on global temperature trends since the Industrial Revolution.[156]

    Climate change feedbacks
    Main articles: Climate change feedbacks and Climate sensitivity

    Sea ice reflects 50% to 70% of incoming sunlight, while the ocean, being darker, reflects only 6%. As an area of sea ice melts and exposes more ocean, more heat is absorbed by the ocean, raising temperatures that melt still more ice. This is a positive feedback process.[158]
    The climate system’s response to an initial forcing is shaped by feedbacks, which either amplify or dampen the change. Self-reinforcing or positive feedbacks increase the response,[159] while balancing or negative feedbacks reduce it.[160] The main reinforcing feedbacks are the water-vapour feedback, the ice–albedo feedback, and the net cloud feedback.[161][162] The primary balancing mechanism is radiative cooling, as Earth’s surface gives off more heat to space in response to rising temperature.[163] In addition to temperature feedbacks, there are feedbacks in the carbon cycle, such as the fertilizing effect of CO2 on plant growth.[164] Feedbacks are expected to trend in a positive direction as greenhouse gas emissions continue, raising climate sensitivity.[165]

    These feedback processes alter the pace of global warming. For instance, warmer air can hold more moisture in the form of water vapour, which is itself a potent greenhouse gas.[161] Warmer air can also make clouds higher and thinner, and therefore more insulating, increasing climate warming.[166] The reduction of snow cover and sea ice in the Arctic is another major feedback, this reduces the reflectivity of the Earth’s surface in the region and accelerates Arctic warming.[167][168] This additional warming also contributes to permafrost thawing, which releases methane and CO2 into the atmosphere.[169]

    Around half of human-caused CO2 emissions have been absorbed by land plants and by the oceans.[170] This fraction is not static and if future CO2 emissions decrease, the Earth will be able to absorb up to around 70%. If they increase substantially, it’ll still absorb more carbon than now, but the overall fraction will decrease to below 40%.[171] This is because climate change increases droughts and heat waves that eventually inhibit plant growth on land, and soils will release more carbon from dead plants when they are warmer.[172][173] The rate at which oceans absorb atmospheric carbon will be lowered as they become more acidic and experience changes in thermohaline circulation and phytoplankton distribution.[174][175][85] Uncertainty over feedbacks, particularly cloud cover,[176] is the major reason why different climate models project different magnitudes of warming for a given amount of emissions.[177]

    Modelling
    Further information: Climate model and Climate change scenario

    Energy flows between space, the atmosphere, and Earth’s surface. Most sunlight passes through the atmosphere to heat the Earth’s surface, then greenhouse gases absorb most of the heat the Earth radiates in response. Adding to greenhouse gases increases this insulating effect, causing an energy imbalance that heats the planet up.
    A climate model is a representation of the physical, chemical and biological processes that affect the climate system.[178] Models include natural processes like changes in the Earth’s orbit, historical changes in the Sun’s activity, and volcanic forcing.[179] Models are used to estimate the degree of warming future emissions will cause when accounting for the strength of climate feedbacks.[180][181] Models also predict the circulation of the oceans, the annual cycle of the seasons, and the flows of carbon between the land surface and the atmosphere.[182]

    The physical realism of models is tested by examining their ability to simulate current or past climates.[183] Past models have underestimated the rate of Arctic shrinkage[184] and underestimated the rate of precipitation increase.[185] Sea level rise since 1990 was underestimated in older models, but more recent models agree well with observations.[186] The 2017 United States-published National Climate Assessment notes that “climate models may still be underestimating or missing relevant feedback processes”.[187] Additionally, climate models may be unable to adequately predict short-term regional climatic shifts.[188]

    A subset of climate models add societal factors to a physical climate model. These models simulate how population, economic growth, and energy use affect—and interact with—the physical climate. With this information, these models can produce scenarios of future greenhouse gas emissions. This is then used as input for physical climate models and carbon cycle models to predict how atmospheric concentrations of greenhouse gases might change.[189][190] Depending on the socioeconomic scenario and the mitigation scenario, models produce atmospheric CO2 concentrations that range widely between 380 and 1400 ppm.[191]

    Impacts
    Main article: Effects of climate change

    In virtually all countries and territories around the world, scientists in the field of extreme event attribution have concluded that human-caused global warming has increased the number of days of extreme heat events over long-term norms.[192]
    Environmental effects
    Further information: Effects of climate change on oceans and Effects of climate change on the water cycle
    The environmental effects of climate change are broad and far-reaching, affecting oceans, ice, and weather. Changes may occur gradually or rapidly. Evidence for these effects comes from studying climate change in the past, from modelling, and from modern observations.[193] Since the 1950s, droughts and heat waves have appeared simultaneously with increasing frequency.[194] Extremely wet or dry events within the monsoon period have increased in India and East Asia.[195] Monsoonal precipitation over the Northern Hemisphere has increased since 1980.[196] The rainfall rate and intensity of hurricanes and typhoons is likely increasing,[197] and the geographic range likely expanding poleward in response to climate warming.[198] The frequency of tropical cyclones has not increased as a result of climate change.[199]

    Historical sea level reconstruction and projections up to 2100 published in 2017 by the U.S. Global Change Research Program[200]
    Global sea level is rising as a consequence of thermal expansion and the melting of glaciers and ice sheets. Sea level rise has increased over time, reaching 4.8 cm per decade between 2014 and 2023.[201] Over the 21st century, the IPCC projects 32–62 cm of sea level rise under a low emission scenario, 44–76 cm under an intermediate one and 65–101 cm under a very high emission scenario.[202] Marine ice sheet instability processes in Antarctica may add substantially to these values,[203] including the possibility of a 2-meter sea level rise by 2100 under high emissions.[204]

    Climate change has led to decades of shrinking and thinning of the Arctic sea ice.[205] While ice-free summers are expected to be rare at 1.5 °C degrees of warming, they are set to occur once every three to ten years at a warming level of 2 °C.[206] Higher atmospheric CO2 concentrations cause more CO2 to dissolve in the oceans, which is making them more acidic.[207] Because oxygen is less soluble in warmer water,[208] its concentrations in the ocean are decreasing, and dead zones are expanding.[209]

    Tipping points and long-term impacts
    Main article: Tipping points in the climate system

    Different levels of global warming may cause different parts of Earth’s climate system to reach tipping points that cause transitions to different states.[210][211]
    Greater degrees of global warming increase the risk of passing through ‘tipping points’ –thresholds beyond which certain major impacts can no longer be avoided even if temperatures return to their previous state.[212][213] For instance, the Greenland ice sheet is already melting, but if global warming reaches levels between 1.7 °C and 2.3 °C, its melting will continue until it fully disappears. If the warming is later reduced to 1.5 °C or less, it will still lose a lot more ice than if the warming was never allowed to reach the threshold in the first place.[214] While the ice sheets would melt over millennia, other tipping points would occur faster and give societies less time to respond. The collapse of major ocean currents like the Atlantic meridional overturning circulation (AMOC), and irreversible damage to key ecosystems like the Amazon rainforest and coral reefs can unfold in a matter of decades.[211] The collapse of the AMOC would be a severe climate catastrophe, resulting in a cooling of the Northern Hemisphere.[215]

    The long-term effects of climate change on oceans include further ice melt, ocean warming, sea level rise, ocean acidification and ocean deoxygenation.[216] The timescale of long-term impacts are centuries to millennia due to CO2’s long atmospheric lifetime.[217] The result is an estimated total sea level rise of 2.3 metres per degree Celsius (4.2 ft/°F) after 2000 years.[218] Oceanic CO2 uptake is slow enough that ocean acidification will also continue for hundreds to thousands of years.[219] Deep oceans (below 2,000 metres (6,600 ft)) are also already committed to losing over 10% of their dissolved oxygen by the warming which occurred to date.[220] Further, the West Antarctic ice sheet appears committed to practically irreversible melting, which would increase the sea levels by at least 3.3 m (10 ft 10 in) over approximately 2000 years.[211][221][222]

    Nature and wildlife
    Further information: Effects of climate change on oceans and Effects of climate change on biomes
    Recent warming has driven many terrestrial and freshwater species poleward and towards higher altitudes.[223] For instance, the range of hundreds of North American birds has shifted northward at an average rate of 1.5 km/year over the past 55 years.[224] Higher atmospheric CO2 levels and an extended growing season have resulted in global greening. However, heatwaves and drought have reduced ecosystem productivity in some regions. The future balance of these opposing effects is unclear.[225] A related phenomenon driven by climate change is woody plant encroachment, affecting up to 500 million hectares globally.[226] Climate change has contributed to the expansion of drier climate zones, such as the expansion of deserts in the subtropics.[227] The size and speed of global warming is making abrupt changes in ecosystems more likely.[228] Overall, it is expected that climate change will result in the extinction of many species.[229]

    The oceans have heated more slowly than the land, but plants and animals in the ocean have migrated towards the colder poles faster than species on land.[230] Just as on land, heat waves in the ocean occur more frequently due to climate change, harming a wide range of organisms such as corals, kelp, and seabirds.[231] Ocean acidification makes it harder for marine calcifying organisms such as mussels, barnacles and corals to produce shells and skeletons; and heatwaves have bleached coral reefs.[232] Harmful algal blooms enhanced by climate change and eutrophication lower oxygen levels, disrupt food webs and cause great loss of marine life.[233] Coastal ecosystems are under particular stress. Almost half of global wetlands have disappeared due to climate change and other human impacts.[234] Plants have come under increased stress from damage by insects.[235]

    Climate change impacts on the environment
    Underwater photograph of branching coral that is bleached white
    Ecological collapse. Coral bleaching from thermal stress has damaged the Great Barrier Reef and threatens coral reefs worldwide.[236]

    Photograph of evening in a valley settlement. The skyline in the hills beyond is lit up red from the fires.
    Extreme weather. Drought and high temperatures worsened the 2020 bushfires in Australia.[237]

    The green landscape is interrupted by a huge muddy scar where the ground has subsided.
    Arctic warming. Permafrost thaws undermine infrastructure and release methane, a greenhouse gas.[169]

    An emaciated polar bear stands atop the remains of a melting ice floe.
    Habitat destruction. Many arctic animals rely on sea ice, which has been disappearing in a warming Arctic.[238]

    Photograph of a large area of forest. The green trees are interspersed with large patches of damaged or dead trees turning purple-brown and light red.
    Pest propagation. Mild winters allow more pine beetles to survive to kill large swaths of forest.[239]
    Humans
    Main article: Effects of climate change

    Extreme weather will be progressively more common as the Earth warms.[240]
    The effects of climate change are impacting humans everywhere in the world.[241] Impacts can be observed on all continents and ocean regions,[242] with low-latitude, less developed areas facing the greatest risk.[243] Continued warming has potentially “severe, pervasive and irreversible impacts” for people and ecosystems.[244] The risks are unevenly distributed, but are generally greater for disadvantaged people in developing and developed countries.[245]

    Health and food
    Main articles: Effects of climate change on agriculture § Global food security and undernutrition, and Effects of climate change on human health
    The World Health Organization calls climate change one of the biggest threats to global health in the 21st century.[15] Scientists have warned about the irreversible harms it poses.[246] Extreme weather events affect public health, and food and water security.[247][248][249] Temperature extremes lead to increased illness and death.[247][248] Climate change increases the intensity and frequency of extreme weather events.[248][249] It can affect transmission of infectious diseases, such as dengue fever and malaria.[246][247] According to the World Economic Forum, 14.5 million more deaths are expected due to climate change by 2050.[250] 30% of the global population currently live in areas where extreme heat and humidity are already associated with excess deaths.[251][252] By 2100, 50% to 75% of the global population would live in such areas.[251][253]

    While total crop yields have been increasing in the past 50 years due to agricultural improvements, climate change has already decreased the rate of yield growth.[249] Fisheries have been negatively affected in multiple regions.[249] While agricultural productivity has been positively affected in some high latitude areas, mid- and low-latitude areas have been negatively affected.[249] Climate change affects land suitability for many crops, with suitable areas for given crops usually moving to higher latitudes and altitudes.[254] According to the World Economic Forum, an increase in drought in certain regions could cause 3.2 million deaths from malnutrition by 2050 and stunting in children.[255] With 2 °C warming, global livestock headcounts could decline by 7–10% by 2050, as less animal feed will be available.[256] If the emissions continue to increase for the rest of century, then over 9 million climate-related deaths would occur annually by 2100.[257]

    Economics, livelihoods and inequality
    Further information: Economic analysis of climate change and Climate security
    Economic damages due to climate change may be severe and there is a chance of disastrous consequences.[258] Severe impacts are expected in South-East Asia and sub-Saharan Africa, where most of the local inhabitants are dependent upon natural and agricultural resources.[259][260] Heat stress can prevent outdoor labourers from working. If warming reaches 4 °C then labour capacity in those regions could be reduced by 30 to 50%.[261] The World Bank estimates that between 2016 and 2030, climate change could drive over 120 million people into extreme poverty without adaptation.[262]

    Inequalities based on wealth and social status have worsened due to climate change.[263] Major difficulties in mitigating, adapting to, and recovering from climate shocks are faced by marginalized people who have less control over resources.[264][259] Indigenous people, who are subsistent on their land and ecosystems, will face endangerment to their wellness and lifestyles due to climate change.[265] An expert elicitation concluded that the role of climate change in armed conflict has been small compared to factors such as socio-economic inequality and state capabilities.[266]

    While women are not inherently more at risk from climate change and shocks, limits on women’s resources and discriminatory gender norms constrain their adaptive capacity and resilience.[267] For example, women’s work burdens, including hours worked in agriculture, tend to decline less than men’s during climate shocks such as heat stress.[267]

    Climate change threatens the sports economy by disrupting seasons, damaging infrastructure and reducing fan engagement, with the most immediate risks seen in football, winter sports and outdoor events.[268]

    Climate migration
    Main article: Climate migration
    Low-lying islands and coastal communities are threatened by sea level rise, which makes urban flooding more common. Sometimes, land is permanently lost to the sea.[269] This could lead to statelessness for people in island nations, such as the Maldives and Tuvalu.[270] In some regions, the rise in temperature and humidity may be too severe for humans to adapt to.[271] With worst-case climate change, models project that areas almost one-third of humanity live in might become Sahara-like uninhabitable and extremely hot climates.[272]

    These factors can drive climate or environmental migration, within and between countries.[273] More people are expected to be displaced because of sea level rise, extreme weather and conflict from increased competition over natural resources. Climate change may also increase vulnerability, leading to “trapped populations” who are not able to move due to a lack of resources.[274]

    Climate change impacts on people
    Environmental migration. Sparser rainfall leads to desertification that harms agriculture and can displace populations. Shown: Telly, Mali (2008).
    Environmental migration. Sparser rainfall leads to desertification that harms agriculture and can displace populations. Shown: Telly, Mali (2008).[275]

    Agricultural changes. Droughts, rising temperatures, and extreme weather negatively impact agriculture. Shown: Texas, US (2013).
    Agricultural changes. Droughts, rising temperatures, and extreme weather negatively impact agriculture. Shown: Texas, US (2013).[276]

    Tidal flooding. Sea-level rise increases flooding in low-lying coastal regions. Shown: Venice, Italy (2004).
    Tidal flooding. Sea-level rise increases flooding in low-lying coastal regions. Shown: Venice, Italy (2004).[277]

    Storm intensification. Bangladesh after Cyclone Sidr (2007) is an example of catastrophic flooding from increased rainfall.
    Storm intensification. Bangladesh after Cyclone Sidr (2007) is an example of catastrophic flooding from increased rainfall.[278]

    Heat wave intensification. Events like the 2022 Southern Cone heat wave are becoming more common.
    Heat wave intensification. Events like the 2022 Southern Cone heat wave are becoming more common.[279]
    Reducing and recapturing emissions
    Further information: Climate change mitigation

    Global greenhouse gas emission scenarios, based on policies and pledges as of November 2021
    Climate change can be mitigated by reducing the rate at which greenhouse gases are emitted into the atmosphere, and by increasing the rate at which carbon dioxide is removed from the atmosphere.[280] To limit global warming to less than 2 °C global greenhouse gas emissions need to be net-zero by 2070.[281] This requires far-reaching, systemic changes on an unprecedented scale in energy, land, cities, transport, buildings, and industry.[282]

    The United Nations Environment Programme estimates that countries need to triple their pledges under the Paris Agreement within the next decade to limit global warming to 2 °C.[283] With pledges made under the Paris Agreement as of 2024, there would be a 66% chance that global warming is kept under 2.8 °C by the end of the century (range: 1.9–3.7 °C, depending on exact implementation and technological progress). When only considering current policies, this raises to 3.1 °C.[284] Globally, limiting warming to 2 °C may result in higher economic benefits than economic costs.[285]

    Although there is no single pathway to limit global warming to 2 °C,[286] most scenarios and strategies see a major increase in the use of renewable energy in combination with increased energy efficiency measures to generate the needed greenhouse gas reductions.[287] To reduce pressures on ecosystems and enhance their carbon sequestration capabilities, changes would also be necessary in agriculture and forestry,[288] such as preventing deforestation and restoring natural ecosystems by reforestation.[289]

    Other approaches to mitigating climate change have a higher level of risk. Scenarios that limit global warming to 1.5 °C typically project the large-scale use of carbon dioxide removal methods over the 21st century.[290] There are concerns, though, about over-reliance on these technologies, and environmental impacts.[291]

    Solar radiation modification (SRM) is a proposal for reducing global warming by reflecting some sunlight away from Earth and back into space. Because it does not reduce greenhouse gas concentrations, it would not address ocean acidification[292] and is not considered mitigation.[293] SRM should be considered only as a supplement to mitigation, not a replacement for it,[294] due to risks such as rapid warming if it were abruptly stopped and not restarted.[295] The most-studied approach is stratospheric aerosol injection.[296] SRM could reduce global warming and some of its impacts, though imperfectly.[297] It poses environmental risks, such as changes to rainfall patterns,[298] as well as political challenges, such as who would decide whether to use it.[296]

    Clean energy
    Main articles: Sustainable energy and Sustainable transport

    Coal, oil, and natural gas remain the primary global energy sources even as renewables have begun rapidly increasing.[299][300]

    Wind and solar power, Germany
    Renewable energy is key to limiting climate change.[301] For decades, fossil fuels have accounted for roughly 80% of the world’s energy use.[302] The remaining share has been split between nuclear power and renewables (including hydropower, bioenergy, wind and solar power and geothermal energy).[303] Fossil fuel use is expected to peak in absolute terms prior to 2030 and then to decline, with coal use experiencing the sharpest reductions.[304] Renewables represented 86% of all new electricity generation installed in 2023.[305] Other forms of clean energy, such as nuclear and hydropower, currently have a larger share of the energy supply. However, their future growth forecasts appear limited in comparison.[306]

    While solar panels and onshore wind are now among the cheapest forms of adding new power generation capacity in many locations,[307] green energy policies are needed to achieve a rapid transition from fossil fuels to renewables.[308] To achieve carbon neutrality by 2050, renewable energy would become the dominant form of electricity generation, rising to 85% or more by 2050 in some scenarios. Investment in coal would be eliminated and coal use nearly phased out by 2050.[309][310]

    Electricity generated from renewable sources would also need to become the main energy source for heating and transport.[311] Transport can switch away from internal combustion engine vehicles and towards electric vehicles, public transit, and active transport (cycling and walking).[312][313] For shipping and flying, low-carbon fuels would reduce emissions.[312] Heating could be increasingly decarbonized with technologies like heat pumps.[314]

    There are obstacles to the continued rapid growth of clean energy, including renewables.[315] Wind and solar produce energy intermittently and with seasonal variability. Traditionally, hydro dams with reservoirs and fossil fuel power plants have been used when variable energy production is low. Going forward, battery storage can be expanded, energy demand and supply can be matched, and long-distance transmission can smooth variability of renewable outputs.[301] Bioenergy is often not carbon-neutral and may have negative consequences for food security.[316] The growth of nuclear power is constrained by controversy around radioactive waste, nuclear weapon proliferation, and accidents.[317][318] Hydropower growth is limited by the fact that the best sites have been developed, and new projects are confronting increased social and environmental concerns.[319]

    Low-carbon energy improves human health by minimizing climate change as well as reducing air pollution deaths,[320] which were estimated at 7 million annually in 2016.[321] Meeting the Paris Agreement goals that limit warming to a 2 °C increase could save about a million of those lives per year by 2050, whereas limiting global warming to 1.5 °C could save millions and simultaneously increase energy security and reduce poverty.[322] Improving air quality also has economic benefits which may be larger than mitigation costs.[323]

    Energy conservation
    Main articles: Efficient energy use and Energy conservation
    Reducing energy demand is another major aspect of reducing emissions.[324] If less energy is needed, there is more flexibility for clean energy development. It also makes it easier to manage the electricity grid, and minimizes carbon-intensive infrastructure development.[325] Major increases in energy efficiency investment will be required to achieve climate goals, comparable to the level of investment in renewable energy.[326] Several COVID-19 related changes in energy use patterns, energy efficiency investments, and funding have made forecasts for this decade more difficult and uncertain.[327]

    Strategies to reduce energy demand vary by sector. In the transport sector, passengers and freight can switch to more efficient travel modes, such as buses and trains, or use electric vehicles.[328] Industrial strategies to reduce energy demand include improving heating systems and motors, designing less energy-intensive products, and increasing product lifetimes.[329] In the building sector the focus is on better design of new buildings, and higher levels of energy efficiency in retrofitting.[330] The use of technologies like heat pumps can also increase building energy efficiency.[331]

    Agriculture and industry
    See also: Sustainable agriculture and Green industrial policy

    Taking into account direct and indirect emissions, industry is the sector with the highest share of global emissions. Data as of 2019 from the IPCC.
    Agriculture and forestry face a triple challenge of limiting greenhouse gas emissions, preventing the further conversion of forests to agricultural land, and meeting increases in world food demand.[332] A set of actions could reduce agriculture and forestry-based emissions by two-thirds from 2010 levels. These include reducing growth in demand for food and other agricultural products, increasing land productivity, protecting and restoring forests, and reducing greenhouse gas emissions from agricultural production.[333]

    On the demand side, a key component of reducing emissions is shifting people towards plant-based diets.[334] Eliminating the production of livestock for meat and dairy would eliminate about three-quarters of all emissions from agriculture and other land use.[335] Livestock also occupy 37% of ice-free land area on Earth and consume feed from the 12% of land area used for crops, driving deforestation and land degradation.[336]

    Steel and cement production are responsible for about 13% of industrial CO2 emissions. In these industries, carbon-intensive materials such as coke and lime play an integral role in the production, so that reducing CO2 emissions requires research into alternative chemistries.[337] Where energy production or CO2-intensive heavy industries continue to produce waste CO2, technology can sometimes be used to capture and store most of the gas instead of releasing it to the atmosphere.[338] This technology, carbon capture and storage (CCS), could have a critical but limited role in reducing emissions.[338] It is relatively expensive[339] and has been deployed only to an extent that removes around 0.1% of annual greenhouse gas emissions.[338]

    Carbon dioxide removal
    Main articles: Carbon dioxide removal and Carbon sequestration

    Most CO2 emissions have been absorbed by carbon sinks, including plant growth, soil uptake, and ocean uptake (2020 Global Carbon Budget).
    Natural carbon sinks can be enhanced to sequester significantly larger amounts of CO2 beyond naturally occurring levels.[340] Reforestation and afforestation (planting forests where there were none before) are among the most mature sequestration techniques, although the latter raises food security concerns.[341] Farmers can promote sequestration of carbon in soils through practices such as use of winter cover crops, reducing the intensity and frequency of tillage, and using compost and manure as soil amendments.[342] Forest and landscape restoration yields many benefits for the climate, including greenhouse gas emissions sequestration and reduction.[133] Restoration/recreation of coastal wetlands, prairie plots and seagrass meadows increases the uptake of carbon into organic matter.[343][344] When carbon is sequestered in soils and in organic matter such as trees, there is a risk of the carbon being re-released into the atmosphere later through changes in land use, fire, or other changes in ecosystems.[345]

    The use of bioenergy in conjunction with carbon capture and storage (BECCS) can result in net negative emissions as CO2 is drawn from the atmosphere.[346] It remains highly uncertain whether carbon dioxide removal techniques will be able to play a large role in limiting warming to 1.5 °C. Policy decisions that rely on carbon dioxide removal increase the risk of global warming rising beyond international goals.[347]

    Adaptation
    Main article: Climate change adaptation
    Adaptation is “the process of adjustment to current or expected changes in climate and its effects”.[348]: 5  Without additional mitigation, adaptation cannot avert the risk of “severe, widespread and irreversible” impacts.[349] More severe climate change requires more transformative adaptation, which can be prohibitively expensive.[350] The capacity and potential for humans to adapt is unevenly distributed across different regions and populations, and developing countries generally have less.[351] The first two decades of the 21st century saw an increase in adaptive capacity in most low- and middle-income countries with improved access to basic sanitation and electricity, but progress is slow. Many countries have implemented adaptation policies. However, there is a considerable gap between necessary and available finance.[352]

    Adaptation to sea level rise consists of avoiding at-risk areas, learning to live with increased flooding, and building flood controls. If that fails, managed retreat may be needed.[353] There are economic barriers for tackling dangerous heat impact. Avoiding strenuous work or having air conditioning is not possible for everybody.[354] In agriculture, adaptation options include a switch to more sustainable diets, diversification, erosion control, and genetic improvements for increased tolerance to a changing climate.[355] Insurance allows for risk-sharing, but is often difficult to get for people on lower incomes.[356] Education, migration and early warning systems can reduce climate vulnerability.[357] Planting mangroves or encouraging other coastal vegetation can buffer storms.[358][359]

    Ecosystems adapt to climate change, a process that can be supported by human intervention. By increasing connectivity between ecosystems, species can migrate to more favourable climate conditions. Species can also be introduced to areas acquiring a favourable climate. Protection and restoration of natural and semi-natural areas helps build resilience, making it easier for ecosystems to adapt. Many of the actions that promote adaptation in ecosystems, also help humans adapt via ecosystem-based adaptation. For instance, restoration of natural fire regimes makes catastrophic fires less likely, and reduces human exposure. Giving rivers more space allows for more water storage in the natural system, reducing flood risk. Restored forest acts as a carbon sink, but planting trees in unsuitable regions can exacerbate climate impacts.[360]

    There are synergies but also trade-offs between adaptation and mitigation.[361] An example for synergy is increased food productivity, which has large benefits for both adaptation and mitigation.[362] An example of a trade-off is that increased use of air conditioning allows people to better cope with heat, but increases energy demand. Another trade-off example is that more compact urban development may reduce emissions from transport and construction, but may also increase the urban heat island effect, exposing people to heat-related health risks.[363]

    Examples of adaptation methods
    Mangrove planting and other habitat conservation can reduce coastal flooding.
    Mangrove planting and other habitat conservation can reduce coastal flooding.

    Seawalls to protect against storm surge worsened by sea level rise
    Seawalls to protect against storm surge worsened by sea level rise

    Green roofs to provide cooling in cities
    Green roofs to provide cooling in cities

    Selective breeding for drought-resistant crops
    Selective breeding for drought-resistant crops
    Policies and politics
    See also: Politics of climate change and Climate change mitigation § Policies
    Map
    Wikimedia | © OpenStreetMap
    The Climate Change Performance Index ranks countries by greenhouse gas emissions (40% of score), renewable energy (20%), energy use (20%), and climate policy (20%).
    High
    Medium
    Low
    Very low
    Countries that are most vulnerable to climate change have typically been responsible for a small share of global emissions. This raises questions about justice and fairness.[364] Limiting global warming makes it much easier to achieve the UN’s Sustainable Development Goals, such as eradicating poverty and reducing inequalities. The connection is recognized in Sustainable Development Goal 13 which is to “take urgent action to combat climate change and its impacts”.[365] The goals on food, clean water and ecosystem protection have synergies with climate mitigation.[366]

    The geopolitics of climate change is complex. It has often been framed as a free-rider problem, in which all countries benefit from mitigation done by other countries, but individual countries would lose from switching to a low-carbon economy themselves. Sometimes mitigation also has localized benefits though. For instance, the benefits of a coal phase-out to public health and local environments exceed the costs in almost all regions.[367] Furthermore, net importers of fossil fuels win economically from switching to clean energy, causing net exporters to face stranded assets: fossil fuels they cannot sell.[368]

    Policy options
    Further information: Climate policy
    A wide range of policies, regulations, and laws are being used to reduce emissions. As of 2019, carbon pricing covers about 20% of global greenhouse gas emissions.[369] Carbon can be priced with carbon taxes and emissions trading systems.[370] Direct global fossil fuel subsidies reached $319 billion in 2017, and $5.2 trillion when indirect costs such as air pollution are priced in.[371] Ending these can cause a 28% reduction in global carbon emissions and a 46% reduction in air pollution deaths.[372] Money saved on fossil subsidies could be used to support the transition to clean energy instead.[373] More direct methods to reduce greenhouse gases include vehicle efficiency standards, renewable fuel standards, and air pollution regulations on heavy industry.[374] Several countries require utilities to increase the share of renewables in power production.[375] An Open Coalition on Compliance Carbon Markets with the aim of creating a global cap and trade system was established at COP30 (2025). According to some calculations it can increase emissions reduction seven-fold over current policies, deliver $200 billion per year for clean-energy and social programs and even close the gap between current emissions trajectory and the goals of the Paris agreement.[376][377][378]

    Climate justice
    Policy designed through the lens of climate justice tries to address human rights issues and social inequality. According to proponents of climate justice, the costs of climate adaptation should be paid by those most responsible for climate change, while the beneficiaries of payments should be those suffering impacts. One way this can be addressed in practice is to have wealthy nations pay poorer countries to adapt.[379]

    Oxfam found that in 2023 the wealthiest 10% of people were responsible for 50% of global emissions, while the bottom 50% were responsible for just 8%.[380] Production of emissions is another way to look at responsibility: under that approach, the top 21 fossil fuel companies would owe cumulative climate reparations of $5.4 trillion over the period 2025–2050.[381] To achieve a just transition, people working in the fossil fuel sector would also need other jobs, and their communities would need investments.[382]

    International climate agreements
    Further information: United Nations Framework Convention on Climate Change

    Since 2000, rising CO2 emissions in China and the rest of world have surpassed the output of the United States and Europe.[383][384]

    Per person, the United States generates CO2 at a far faster rate than other primary regions.[383][385]
    Nearly all countries in the world are parties to the 1994 United Nations Framework Convention on Climate Change (UNFCCC).[386] The goal of the UNFCCC is to prevent dangerous human interference with the climate system.[387] As stated in the convention, this requires that greenhouse gas concentrations are stabilized in the atmosphere at a level where ecosystems can adapt naturally to climate change, food production is not threatened, and economic development can be sustained.[388] The UNFCCC does not itself restrict emissions but rather provides a framework for protocols that do. Global emissions have risen since the UNFCCC was signed.[389] Its yearly conferences are the stage of global negotiations.[390]

    The 1997 Kyoto Protocol extended the UNFCCC and included legally binding commitments for most developed countries to limit their emissions.[391] During the negotiations, the G77 (representing developing countries) pushed for a mandate requiring developed countries to “[take] the lead” in reducing their emissions,[392] since developed countries contributed most to the accumulation of greenhouse gases in the atmosphere. Per-capita emissions were also still relatively low in developing countries and developing countries would need to emit more to meet their development needs.[393]

    The 2009 Copenhagen Accord has been widely portrayed as disappointing because of its low goals, and was rejected by poorer nations including the G77.[394] Associated parties aimed to limit the global temperature rise to below 2 °C.[395] The accord set the goal of sending $100 billion per year to developing countries for mitigation and adaptation by 2020, and proposed the founding of the Green Climate Fund.[396] As of 2020, only $83.3 billion were delivered. Only in 2023 the target is expected to be achieved.[397]

    In 2015 all UN countries negotiated the Paris Agreement, which aims to keep global warming well below 2.0 °C and contains an aspirational goal of keeping warming under 1.5 °C.[398] The agreement replaced the Kyoto Protocol. Unlike Kyoto, no binding emission targets were set in the Paris Agreement. Instead, a set of procedures was made binding. Countries have to regularly set ever more ambitious goals and reevaluate these goals every five years.[399] The Paris Agreement restated that developing countries must be financially supported.[400] As of August 2026, 194 states and the European Union have acceded to or ratified the agreement.[401]

    The 1987 Montreal Protocol, an international agreement to phase out production of ozone-depleting gases, has had benefits for climate change mitigation.[402] Several ozone-depleting gases like chlorofluorocarbons are powerful greenhouse gases, so banning their production and usage may have avoided a temperature rise of 0.5 °C–1.0 °C,[403] as well as additional warming by preventing damage to vegetation from ultraviolet radiation.[404] It is estimated that the agreement has been more effective at curbing greenhouse gas emissions than the Kyoto Protocol specifically designed to do so.[405] The most recent amendment to the Montreal Protocol, the 2016 Kigali Amendment, committed to reducing the emissions of hydrofluorocarbons, which served as a replacement for banned ozone-depleting gases and are also potent greenhouse gases.[406] Should countries comply with the amendment, a warming of 0.3 °C–0.5 °C is estimated to be avoided.[407]

    National responses

    Annual CO2 emissions by region. This measures fossil fuel and industry emissions. Land use change is not included.[408]
    In 2019, the United Kingdom parliament became the first national government to declare a climate emergency.[409] Other countries and jurisdictions followed suit.[410] That same year, the European Parliament declared a “climate and environmental emergency”.[411] The European Commission presented its European Green Deal with the goal of making the EU carbon-neutral by 2050.[412] In 2021, the European Commission released its “Fit for 55” legislation package, which contains guidelines for the car industry; all new cars on the European market must be zero-emission vehicles from 2035.[413]

    Major countries in Asia have made similar pledges: South Korea and Japan have committed to become carbon-neutral by 2050, and China by 2060.[414] While India has strong incentives for renewables, it also plans a significant expansion of coal in the country.[415] Vietnam is among very few coal-dependent, fast-developing countries that pledged to phase out unabated coal power by the 2040s or as soon as possible thereafter.[416]

    As of 2021, based on information from 48 national climate plans representing 40% of the parties to the Paris Agreement, the estimated total greenhouse gas emissions were 0.5% lower than 2010 levels. This is nowhere near the 45% reduction goal to limit global warming to 1.5 °C or the 25% reduction goal to limit it to 2 °C.[417][needs update]

    Society and culture
    Denial and misinformation
    Further information: Climate change denial and Fossil fuels lobby

    Data has been cherry picked from short periods to falsely assert that global temperatures are not rising. Blue trendlines show short periods that mask longer-term warming trends (red trendlines). Blue rectangle with blue dots shows the so-called global warming hiatus.[418]
    Public debate about climate change has been strongly affected by climate change denial and misinformation, which first emerged in the United States and has since spread to other countries, particularly Canada and Australia. It originated from fossil fuel companies, industry groups, conservative think tanks, and contrarian scientists.[419] Like the tobacco industry, the main strategy of these groups has been to manufacture doubt about climate-change related scientific data and results.[420] People who hold unwarranted doubt about climate change are sometimes called climate change “skeptics”, although “contrarians” or “deniers” are more appropriate terms.[421]

    There are different variants of climate denial: some deny that warming takes place at all, some acknowledge warming but attribute it to natural influences, and some minimize the negative impacts of climate change.[422] Manufacturing uncertainty about the science later developed into a manufactured controversy: creating the belief that there is significant uncertainty about climate change within the scientific community to delay policy changes.[423] Strategies to promote these ideas include criticism of scientific institutions,[424] and questioning the motives of individual scientists.[422] An echo chamber of climate-denying blogs and media has further fomented misunderstanding of climate change.[425]

    Public awareness and opinion
    Further information: Climate communication, Media coverage of climate change, and Public opinion on climate change

    The public substantially underestimates the degree of scientific consensus that humans are causing climate change (2022 data).[426] Studies from 2019 to 2021[427][5][428] found scientific consensus to range from 98.7 to 100%.
    Climate change came to international public attention in the late 1980s.[429] Due to media coverage in the early 1990s, people often confused climate change with other environmental issues like ozone depletion.[430] In popular culture, the climate fiction movie The Day After Tomorrow (2004) and the Al Gore documentary An Inconvenient Truth (2006) focused on climate change.[429]

    Significant regional, gender, age and political differences exist in both public concern for, and understanding of, climate change. More highly educated people, and in some countries, women and younger people, were more likely to see climate change as a serious threat.[431] College biology textbooks from the 2010s featured less content on climate change compared to those from the preceding decade, with decreasing emphasis on solutions.[432] Partisan gaps also exist in many countries,[433] and countries with high CO2 emissions tend to be less concerned.[434] Views on causes of climate change vary widely between countries.[435] Media coverage linked to protests has had impacts on public sentiment as well as on which aspects of climate change are focused upon.[436] Higher levels of worry are associated with stronger public support for policies that address climate change.[437] Concern has increased over time,[438] and in 2021 a majority of citizens in 30 countries expressed a high level of worry about climate change, or view it as a global emergency.[439] A 2024 survey across 125 countries found that 89% of the global population demanded intensified political action, but systematically underestimated other peoples’ willingness to act.[25][26]

    Climate movement
    Main articles: Climate movement and Climate change litigation
    Climate protests demand that political leaders take action to prevent climate change. They can take the form of public demonstrations, fossil fuel divestment, lawsuits and other activities.[440][441] Prominent demonstrations include the School Strike for Climate. In this initiative, young people across the globe have been protesting since 2018 by skipping school on Fridays, inspired by Swedish activist and then-teenager Greta Thunberg.[442] Mass civil disobedience actions by groups like Extinction Rebellion have protested by disrupting roads and public transport.[443]

    Litigation is increasingly used as a tool to strengthen climate action from public institutions and companies. Activists also initiate lawsuits which target governments and demand that they take ambitious action or enforce existing laws on climate change.[444] Lawsuits against fossil-fuel companies generally seek compensation for loss and damage.[445] In 2025, the UN’s International Court of Justice issued its advisory opinion, saying explicitly that states must act to stop climate change, and if they fail to accomplish that duty, other states can sue them. This obligation includes implementing their commitments in international agreements they are parties to, such as the 2015 Paris Climate Accord.[446][447][448]

    History
    For broader coverage of this topic, see History of climate change science.
    Early discoveries

    Eunice Newton Foote showed carbon dioxide’s heat-capturing effect in 1856, foreseeing its implications for the planet.[449] (Carbon dioxide was called “carbonic acid gas”.)
    Scientists in the 19th century such as Alexander von Humboldt began to foresee the effects of climate change.[450][451][452][453] In the 1820s, Joseph Fourier proposed the greenhouse effect to explain why Earth’s temperature was higher than the Sun’s energy alone could explain. Earth’s atmosphere is transparent to sunlight, so sunlight reaches the surface where it is converted to heat. However, the atmosphere is not transparent to heat radiating from the surface, and captures some of that heat, which in turn warms the planet.[454] In 1856 Eunice Newton Foote demonstrated that the warming effect of the Sun is greater for air with water vapour than for dry air, and that the effect is even greater with carbon dioxide (CO2). In “Circumstances Affecting the Heat of the Sun’s Rays” she concluded that “[a]n atmosphere of that gas would give to our earth a high temperature”.[455][456]

    This 1912 article succinctly describes the greenhouse effect, how burning coal creates carbon dioxide to cause global warming and climate change.[457]
    Starting in 1859,[458] John Tyndall established that nitrogen and oxygen—together totalling 99% of dry air—are transparent to radiated heat. However, water vapour and gases such as methane and carbon dioxide absorb radiated heat and re-radiate that heat into the atmosphere. Tyndall proposed that changes in the concentrations of these gases may have caused climatic changes in the past, including ice ages.[459]

    Svante Arrhenius noted that water vapour in air continuously varied, but the CO2 concentration in air was influenced by long-term geological processes. Warming from increased CO2 levels would increase the amount of water vapour, amplifying warming in a positive feedback loop. In 1896, he published the first climate model of its kind, projecting that halving CO2 levels could have produced a drop in temperature initiating an ice age. Arrhenius calculated the temperature increase expected from doubling CO2 to be around 5–6 °C.[460] Other scientists were initially sceptical and believed that the greenhouse effect was saturated so that adding more CO2 would make no difference, and that the climate would be self-regulating.[461] Beginning in 1938, Guy Stewart Callendar published evidence that climate was warming and CO2 levels were rising,[462] but his calculations met the same objections.[461]

    Development of a scientific consensus
    See also: Scientific consensus on climate change

    Scientific consensus on causation: Academic studies of scientific agreement on human-caused global warming among climate experts (2010–2015) reflect that the level of consensus correlates with expertise in climate science.[463] A 2019 study found scientific consensus to be at 100%,[427] and a 2021 study concluded that consensus exceeded 99%.[5] Another 2021 study found that 98.7% of climate experts indicated that the Earth is getting warmer mostly because of human activity.[428]
    In the 1950s, Gilbert Plass created a detailed computer model that included different atmospheric layers and the infrared spectrum. This model predicted that increasing CO2 levels would cause warming. Around the same time, Hans Suess found evidence that CO2 levels had been rising, and Roger Revelle showed that the oceans would not absorb the increase. The two scientists subsequently helped Charles Keeling to begin a record of continued increase—the “Keeling Curve”[461]—which was part of continued scientific investigation through the 1960s into possible human causation of global warming.[464] Studies such as the National Research Council’s 1979 Charney Report supported the accuracy of climate models that forecast significant warming.[465] Human causation of observed global warming and dangers of unmitigated warming were publicly presented in James Hansen’s 1988 testimony before a US Senate committee.[466][40] The Intergovernmental Panel on Climate Change (IPCC), set up in 1988 to provide formal advice to the world’s governments, spurred interdisciplinary research.[467] As part of the IPCC reports, scientists assess the scientific discussion that takes place in peer-reviewed journal articles.[468]

    There is a nearly unanimous scientific consensus that the climate is warming and that this is caused by human activities.[5] No scientific body of national or international standing disagrees with this view.[469] As of 2019, agreement in recent literature reached over 99%.[427][5] The 2021 IPCC Assessment Report stated that it is “unequivocal” that climate change is caused by humans.[5] Consensus has further developed that action should be taken to protect people against the impacts of climate change. National science academies have called on world leaders to cut global emissions.[470]

    Recent developments
    Extreme event attribution (EEA), also known as attribution science, was developed in the early decades of the 21st century.[471] EEA uses climate models to identify and quantify the role that human-caused climate change plays in the frequency, intensity, duration, and impacts of specific individual extreme weather events.[472][473] Results of attribution studies allow scientists and journalists to make statements such as, “this weather event was made at least n times more likely by human-caused climate change” or “this heatwave was made m degrees hotter than it would have been in a world without global warming” or “this event was effectively impossible without climate change”.[474]

    Greater computing power in the 2000s and conceptual breakthroughs in the early to mid 2010s[475] enabled attribution science to detect the effects of climate change on some events with high confidence.[471] Scientists use attribution methods and climate simulations that have already been peer reviewed, allowing “rapid attribution studies” to be published within a “news cycle” time frame after weather events.[475]

  • Blockchain

    A blockchain is a distributed ledger with growing lists of records (blocks) that are securely linked together via cryptographic hashes.[1][2][3][4] Each block contains a cryptographic hash of the previous block, a timestamp, and transaction data (generally represented as a Merkle tree, where data nodes are represented by leaves). Since each block contains information about the previous block, they effectively form a chain (viz. linked list data structure), with each additional block linking to the ones before it. Consequently, blockchain transactions are resistant to alteration because, once recorded, the data in any given block cannot be changed retroactively without altering all subsequent blocks and obtaining network consensus to accept these changes.

    Blockchains are typically managed by a peer-to-peer (P2P) computer network for use as a public distributed ledger, where nodes collectively adhere to a consensus algorithm protocol to add and validate new transaction blocks. Although blockchain records are not unalterable, since blockchain forks are possible, blockchains may be considered secure by design and exemplify a distributed computing system with high Byzantine fault tolerance.[5]

    A blockchain was created by a person (or group of people) using the name (or pseudonym) Satoshi Nakamoto in 2008 to serve as the public distributed ledger for bitcoin cryptocurrency transactions, based on previous work by Stuart Haber, W. Scott Stornetta, and Dave Bayer.[6] The implementation of the blockchain within bitcoin made it the first digital currency to solve the double-spending problem without the need for a trusted authority or central server. The bitcoin design has inspired other applications[3][2] and blockchains that are readable by the public and are widely used by cryptocurrencies. The blockchain may be considered a type of payment rail.[7]

    Private blockchains have been proposed for business use. Computerworld called the marketing of such privatized blockchains without a proper security model “snake oil”;[8] however, others have argued that permissioned blockchains, if carefully designed, may be more decentralized and therefore more secure in practice than permissionless ones.[4][9]

    History
    Cryptographer David Chaum first proposed a blockchain-like protocol in his 1982 dissertation “Computer Systems Established, Maintained, and Trusted by Mutually Suspicious Groups”.[10][11] Further work on a cryptographically secured chain of blocks was described in 1991 by Stuart Haber and W. Scott Stornetta.[4][12] They wanted to implement a system wherein document timestamps could not be tampered with. In 1992, Haber, Stornetta, and Dave Bayer incorporated Merkle trees into the design, which improved its efficiency by allowing several document certificates to be collected into one block.[4][13] Under their company Surety, their document certificate hashes have been published in The New York Times every week since 1995.[6]

    The first decentralized blockchain was conceptualized by a person (or group of people) known as Satoshi Nakamoto in 2008. Nakamoto improved the design in an important way using a Hashcash-like method to timestamp blocks without requiring them to be signed by a trusted party and introducing a difficulty parameter to stabilize the rate at which blocks are added to the chain.[4] The design was implemented the following year by Nakamoto as a core component of the cryptocurrency bitcoin, where it serves as the public ledger for all transactions on the network.[3]

    In August 2014, the bitcoin blockchain file size, containing records of all transactions that have occurred on the network, reached 20 GB (gigabytes).[14] By 2024, the bitcoin blockchain exceeded 600 GB.[15]

    The words block and chain were used separately in Satoshi Nakamoto’s original paper, but were eventually popularized as a single word, blockchain, by 2016.[16]

    According to Accenture, an application of the diffusion of innovations theory suggests that blockchains attained a 13.5% adoption rate within financial services in 2016, therefore reaching the early adopters’ phase.[17] Industry trade groups joined to create the Global Blockchain Forum in 2016, an initiative of the Chamber of Digital Commerce.

    In May 2018, Gartner found that only 1% of CIOs indicated any kind of blockchain adoption within their organisations, and only 8% of CIOs were in the short-term “planning or [looking at] active experimentation with blockchain”.[18] For the year 2019 Gartner reported 5% of CIOs believed blockchain technology was a ‘game-changer’ for their business.[19]

    Structure and design

    Blockchain formation. The main chain (black) consists of the longest series of blocks from the genesis block (green) to the current block. Orphan blocks (purple) exist outside of the main chain.
    A blockchain is a decentralized, distributed, and often public, digital ledger consisting of records called blocks that are used to record transactions across many computers so that any involved block cannot be altered retroactively, without the alteration of all subsequent blocks.[3][20] This allows the participants to verify and audit transactions independently and relatively inexpensively.[21] A blockchain database is managed autonomously using a peer-to-peer network and a distributed timestamping server. They are authenticated by mass collaboration powered by collective self-interests.[22] Such a design facilitates robust workflow where participants’ uncertainty regarding data security is marginal. The use of a blockchain removes the characteristic of infinite reproducibility from a digital asset. It confirms that each unit of value was transferred only once, solving the long-standing problem of double-spending. A blockchain has been described as a value-exchange protocol.[23]

    Logically, a blockchain can be seen as consisting of several layers:[24]

    infrastructure (hardware)
    networking (node discovery, information propagation[25] and verification)
    consensus (proof of work, proof of stake)
    data (blocks, transactions)
    application (smart contracts/decentralized applications, if applicable)
    Blocks
    Blocks hold batches of valid transactions that are hashed and encoded into a Merkle tree.[3] Each block includes the cryptographic hash of the prior block in the blockchain, linking the two. The linked blocks form a chain.[3] This iterative process confirms the integrity of the previous block, all the way back to the initial block, which is known as the genesis block (Block 0).[26] To assure the integrity of a block and the data contained in it, the block is usually digitally signed.[27]

    Sometimes separate blocks can be produced concurrently, creating a temporary fork. In addition to a secure hash-based history, any blockchain has a specified algorithm for scoring different versions of the history so that one with a higher score can be selected over others. Blocks not selected for inclusion in the chain are called orphan blocks.[26] Peers supporting the database have different versions of the history from time to time. They keep only the highest-scoring version of the database known to them. Whenever a peer receives a higher-scoring version (usually the old version with a single new block added) they extend or overwrite their own database and retransmit the improvement to their peers. There is never an absolute guarantee that any particular entry will remain in the best version of history forever. Blockchains are typically built to add the score of new blocks onto old blocks and are given incentives to extend with new blocks rather than overwrite old blocks. Therefore, the probability of an entry becoming superseded decreases exponentially[28] as more blocks are built on top of it, eventually becoming very low.[3][29]: ch. 08 [30] For example, bitcoin uses a proof-of-work system, where the chain with the most cumulative proof-of-work is considered the valid one by the network. There are a number of methods that can be used to demonstrate a sufficient level of computation. Within a blockchain the computation is carried out redundantly rather than in the traditional segregated and parallel manner.[31]

    Block time
    The block time is the average time it takes for the network to generate one extra block in the blockchain. By the time of block completion, the included data becomes verifiable. In cryptocurrency, this is practically when the transaction takes place, so a shorter block time means faster transactions. Since 2022, the block time for Ethereum is set to exactly 12 seconds, while for bitcoin it is on average 10 minutes.[32]

    Hard forks
    This section is an excerpt from Fork (blockchain) § Hard fork.edit
    A hard fork is a change to the blockchain protocol that is not backward compatible and requires all users to upgrade their software in order to continue participating in the network. In a hard fork, the network splits into two separate versions: one that follows the new rules and one that follows the old rules.

    For example, Ethereum was hard forked in 2016 to “make whole” the investors in The DAO, which had been hacked by exploiting a vulnerability in its code. In this case, the fork resulted in a split creating Ethereum and Ethereum Classic chains. In 2014, the Nxt community[33] was asked to consider a hard fork that would have led to a rollback of the blockchain records to mitigate the effects of a theft of 50 million NXT from a major cryptocurrency exchange. The hard fork proposal was rejected, and some of the funds were recovered after negotiations and ransom payment. Alternatively, to prevent a permanent split, a majority of nodes using the new software may return to the old rules, as was the case with Bitcoin split on 12 March 2013.

    A more recent hard-fork example is of Bitcoin in 2017, which resulted in a split creating Bitcoin Cash.[34] The network split was mainly due to a disagreement in how to increase the transactions per second to accommodate for demand.[35]

    Decentralization
    By storing data across its peer-to-peer network, the blockchain eliminates some risks that come with data being held centrally.[3] The decentralized blockchain may use ad hoc message passing and distributed networking.[36]

    In a so-called “51% attack” a central entity gains control of more than half of a network and can then manipulate that specific blockchain record at will, allowing double-spending.[37]

    Blockchain security methods include the use of public-key cryptography.[38]: 5  A public key (a long, random-looking string of numbers) is an address on the blockchain. Value tokens sent across the network are recorded as belonging to that address. A private key is like a password that gives its owner access to their digital assets or the means to otherwise interact with the various capabilities that blockchains now support. Data stored on the blockchain is generally considered incorruptible.[3]

    Every node in a decentralized system has a copy of the blockchain. Data quality is maintained by massive database replication[39] and computational trust. No centralized “official” copy exists and no user is “trusted” more than any other.[38] Transactions are broadcast to the network using the software. Messages are delivered on a best-effort basis. Early blockchains rely on energy-intensive mining nodes to validate transactions,[26] add them to the block they are building, and then broadcast the completed block to other nodes.[29]: ch. 08  Blockchains use various time-stamping schemes, such as proof-of-work, to serialize changes.[40] Later consensus methods include proof of stake.[26] The growth of a decentralized blockchain is accompanied by the risk of centralization because the computer resources required to process larger amounts of data become more expensive.[41]

    Finality
    Finality is the level of confidence that the well-formed block recently appended to the blockchain will not be revoked in the future (is “finalized”) and thus can be trusted. Most distributed blockchain protocols, whether proof of work or proof of stake, cannot guarantee the finality of a freshly committed block, and instead rely on “probabilistic finality”: as the block goes deeper into a blockchain, it is less likely to be altered or reverted by a newly found consensus.[42]

    Byzantine fault tolerance-based proof-of-stake protocols purport to provide so called “absolute finality”: a randomly chosen validator proposes a block, the rest of validators vote on it, and, if a supermajority decision approves it, the block is irreversibly committed into the blockchain.[42] A modification of this method, an “economic finality”, is used in practical protocols, like the Casper protocol used in Ethereum: validators which sign two different blocks at the same position in the blockchain are subject to “slashing”, where their leveraged stake is forfeited.[42]

    Openness
    Open blockchains are more user-friendly than some traditional ownership records, which, while open to the public, still require physical access to view. Because all early blockchains were permissionless, controversy has arisen over the blockchain definition. An issue in this ongoing debate is whether a private system with verifiers tasked and authorized (permissioned) by a central authority should be considered a blockchain.[43][44][45][46][47] Proponents of permissioned or private chains argue that the term “blockchain” may be applied to any data structure that batches data into time-stamped blocks. These blockchains serve as a distributed version of multiversion concurrency control (MVCC) in databases.[48] Just as MVCC prevents two transactions from concurrently modifying a single object in a database, blockchains prevent two transactions from spending the same single output in a blockchain.[49]: 30–31  Opponents say that permissioned systems resemble traditional corporate databases, not supporting decentralized data verification, and that such systems are not hardened against operator tampering and revision.[43][45] Nikolai Hampton of Computerworld said that “many in-house blockchain solutions will be nothing more than cumbersome databases,” and “without a clear security model, proprietary blockchains should be eyed with suspicion.”[8][50]

    Permissionless (public) blockchain
    An advantage to an open, permissionless, or public, blockchain network is that guarding against bad actors is not required and no access control is needed.[28] This means that applications can be added to the network without the approval or trust of others, using the blockchain as a transport layer.[28]

    Bitcoin and other cryptocurrencies currently secure their blockchain by requiring new entries to include proof of work. To prolong the blockchain, bitcoin uses Hashcash puzzles. While Hashcash was designed in 1997 by Adam Back, the original idea was first proposed by Cynthia Dwork and Moni Naor and Eli Ponyatovski in their 1992 paper “Pricing via Processing or Combatting Junk Mail”.

    In 2016, venture capital investment for blockchain-related projects was weakening in the US but increasing in China.[51] Bitcoin and many other cryptocurrencies use open (public) blockchains. As of April 2018, bitcoin has the highest market capitalization.

    Permissioned (private) blockchain
    See also: Distributed ledger
    Permissioned blockchains use an access control layer to govern who has access to the network.[52] It has been argued that permissioned blockchains can guarantee a certain level of decentralization, if carefully designed, as opposed to permissionless blockchains, which are often centralized in practice.[9]

    Disadvantages of permissioned blockchain
    Nikolai Hampton argued in Computerworld that “There is also no need for a ’51 percent’ attack on a private blockchain, as the private blockchain (most likely) already controls 100 percent of all block creation resources. If you could attack or damage the blockchain creation tools on a private corporate server, you could effectively control 100 percent of their network and alter transactions however you wished.”[8] This has a set of particularly profound adverse implications during a financial crisis or debt crisis such as the 2008 financial crisis, where politically powerful actors may make decisions that favor some groups at the expense of others,[53] and “the bitcoin blockchain is protected by the massive group mining effort. It’s unlikely that any private blockchain will try to protect records using gigawatts of computing power — it’s time-consuming and expensive.”[8] He also said, “Within a private blockchain there is also no ‘race’; there’s no incentive to use more power or discover blocks faster than competitors. This means that many in-house blockchain solutions will be nothing more than cumbersome databases.”[8]

    Blockchain analysis
    The analysis of public blockchains has become increasingly important with the popularity of bitcoin, Ethereum, litecoin and other cryptocurrencies.[54] A blockchain, if it is public, provides access to anyone to observe and analyse the chain data, given the know-how. The process of understanding and accessing the flow of crypto has been an issue for many cryptocurrencies, crypto exchanges and banks.[55][56] The reason for this is accusations of blockchain-enabled cryptocurrencies enabling illicit dark market trading of drugs, weapons, money laundering, etc.[57] A common belief has been that cryptocurrency is private and untraceable, thus leading many actors to use it for illegal purposes. This is changing now that specialised tech companies provide blockchain tracking services, making crypto exchanges, law-enforcement and banks more aware of what is happening with crypto funds and fiat-crypto exchanges. The development, some argue, has led criminals to prioritise the use of new cryptos such as Monero.[58][59][60]

    Standardisation
    In April 2016, Standards Australia submitted a proposal to the International Organization for Standardization to consider developing standards to support blockchain technology. This proposal resulted in the creation of ISO Technical Committee 307, Blockchain and Distributed Ledger Technologies.[61] The technical committee has working groups relating to blockchain terminology, reference architecture, security and privacy, identity, smart contracts, governance and interoperability for blockchain and DLT, as well as standards specific to industry sectors and generic government requirements.[62][non-primary source needed] More than 50 countries are participating in the standardization process together with external liaisons such as the Society for Worldwide Interbank Financial Telecommunication (SWIFT), the European Commission, the International Federation of Surveyors, the International Telecommunication Union (ITU) and the United Nations Economic Commission for Europe (UNECE).[62]

    Many other national standards bodies and open standards bodies are also working on blockchain standards.[63] These include the National Institute of Standards and Technology[64] (NIST), the European Committee for Electrotechnical Standardization[65] (CENELEC), the Institute of Electrical and Electronics Engineers[66] (IEEE), the Organization for the Advancement of Structured Information Standards (OASIS), and some individual participants in the Internet Engineering Task Force[67] (IETF).

    Centralized blockchain
    Although most of blockchain implementation are decentralized and distributed, Oracle launched a centralized blockchain table feature in Oracle 21c database. The Blockchain Table in Oracle 21c database is a centralized blockchain which provide immutable feature. Compared to decentralized blockchains, centralized blockchains normally can provide a higher throughput and lower latency of transactions than consensus-based distributed blockchains.[68][69]

    Types
    Currently, there are at least four types of blockchain networks — public blockchains, private blockchains, consortium blockchains and hybrid blockchains.

    Public blockchains
    A public blockchain has no access restrictions. Anyone with an Internet connection can send transactions to it as well as become a validator (i.e., participate in the execution of a consensus protocol).[70][self-published source?] Usually, such networks offer economic incentives for those who secure them and utilize some type of a proof-of-stake or proof-of-work algorithm.

    Some of the largest, most known public blockchains are the bitcoin blockchain and the Ethereum blockchain.

    Private blockchains
    A private blockchain is permissioned.[52] One cannot join it unless invited by the network administrators. Participant and validator access is restricted. To distinguish between open blockchains and other peer-to-peer decentralized database applications that are not open ad-hoc compute clusters, the terminology Distributed Ledger (DLT) is normally used for private blockchains.

    Hybrid blockchains
    A hybrid blockchain has a combination of centralized and decentralized features.[71] The exact workings of the chain can vary based on which portions of centralization and decentralization are used.

    Sidechains
    A sidechain is a designation for a blockchain ledger that runs in parallel to a primary blockchain.[72][73] Entries from the primary blockchain (where said entries typically represent digital assets) can be linked to and from the sidechain; this allows the sidechain to otherwise operate independently of the primary blockchain (e.g., by using an alternate means of record keeping, alternate consensus algorithm, etc.).[74][better source needed]

    Consortium blockchain
    A consortium blockchain is a type of blockchain that combines elements of both public and private blockchains. In a consortium blockchain, a group of organizations come together to create and operate the blockchain, rather than a single entity. The consortium members jointly manage the blockchain network and are responsible for validating transactions. Consortium blockchains are permissioned, meaning that only certain individuals or organizations are allowed to participate in the network. This allows for greater control over who can access the blockchain and helps to ensure that sensitive information is kept confidential.

    Consortium blockchains are commonly used in industries where multiple organizations need to collaborate on a common goal, such as supply chain management or financial services. One advantage of consortium blockchains is that they can be more efficient and scalable than public blockchains, as the number of nodes required to validate transactions is typically smaller. Additionally, consortium blockchains can provide greater security and reliability than private blockchains, as the consortium members work together to maintain the network. Some examples of consortium blockchains include Quorum and Hyperledger.[75]

    Uses
    The primary use of blockchains is as a distributed ledger for cryptocurrencies such as bitcoin; there were also a few other operational products that had matured from proof of concept by late 2016.[51] As of 2016, some businesses have been testing the technology and conducting low-level implementation to gauge blockchain’s effects on organizational efficiency in their back office.[76]

    In 2019, it was estimated that around $2.9 billion were invested in blockchain technology, which represents an 89% increase from the year prior.[77] A report from to PricewaterhouseCoopers (PwC) in 2021 said that blockchain technology had the potential to generate an annual business value of more than $3 trillion by 2030. PwC surveyed 600 business executives and determined that 84% have at least some exposure to utilizing blockchain technology.[78]

    In 2019, the BBC World Service radio and podcast series 50 Things That Made the Modern Economy identified blockchain as a technology that would have far-reaching consequences for economics and society. The economist and Financial Times journalist and broadcaster Tim Harford discussed why the underlying technology might have much wider applications and the challenges that needed to be overcome.[79] His first broadcast was on 29 June 2019.

    Cryptocurrencies
    Most cryptocurrencies are designed to gradually decrease the production of that currency, placing a cap on the total amount of that currency that will ever be in circulation.[80] Compared with ordinary currencies held by financial institutions or kept as cash on hand, cryptocurrencies can be more difficult for seizure by law enforcement.[80]

    The validity of each cryptocurrency’s coins is provided by a blockchain. A blockchain is a continuously growing list of records, called blocks, which are linked and secured using cryptography.[80][5] Each block typically contains a hash pointer as a link to a previous block,[5] a timestamp and transaction data.[81] By design, blockchains are inherently resistant to modification of the data. It is “an open, distributed ledger that can record transactions between two parties efficiently and in a verifiable and permanent way”.[5] For use as a distributed ledger, a blockchain is typically managed by a peer-to-peer network collectively adhering to a protocol for validating new blocks. Once recorded, the data in any given block cannot be altered retroactively without the alteration of all subsequent blocks, which requires collusion of the network majority.

    Blockchains are secure by design and are an example of a distributed computing system with high Byzantine fault tolerance. Decentralized consensus has therefore been achieved with a blockchain.[82]

    In the context of cryptocurrencies, the blockchain serves as a public ledger for all transactions. Cryptocurrencies use various timestamping schemes to “prove the validity of transactions added to the blockchain ledger without the need for a trusted third party”.[80]

    The first cryptocurrency was Bitcoin, which was first released as open-source software in 2009.

    As cryptocurrencies have gained prominence, several countries have made advancements in their private and commercial law treatment to address legal uncertainties. In the United States, for example, the 2022 amendments to the Uniform Commercial Code (UCC) introduced Article 12, which establishes “controllable electronic records” (CERs) as a new category of personal property. This framework provides legal clarity for the ownership, transfer, and use of cryptocurrencies as CERs, with the concept of “control” serving as a functional equivalent to possession for digital assets. These reforms aim to align legal standards with market practices, reducing title disputes and supporting the integration of cryptocurrencies into commercial transactions.[83]

    Smart contracts
    Main article: Smart contract
    Blockchain-based smart contracts are contracts that can be partially or fully executed or enforced without human interaction.[84] One of the main objectives of a smart contract is automated escrow. A key feature of smart contracts is that they do not need a trusted third party (such as a trustee) to act as an intermediary between contracting entities — the blockchain network executes the contract on its own. This may reduce friction between entities when transferring value and could subsequently open the door to a higher level of transaction automation.[85] An IMF staff discussion from 2018 reported that smart contracts based on blockchain technology might reduce moral hazards and optimize the use of contracts in general, but “no viable smart contract systems have yet emerged.” Due to the lack of widespread use, their legal status was unclear.[86][87]

    Financial services
    According to Reason in 2016, many banks have expressed interest in implementing distributed ledgers for use in banking and are cooperating with companies creating private blockchains.[88] According to a September 2016 IBM study, this implementation was occurring faster than expected.[89]

    Blockchain technology has been proposed as a way to speed up back office settlement systems.[90]

    In 2017, UBS opened a lab dedicated to researching how blockchain technology could be used in financial services.[91]

    In 2017, the German bank Berenberg described blockchain as an “overhyped technology” that has had a large number of “proofs of concept”, but still has major challenges, and very few success stories.[92]

    Blockchain technology has allowed for new forms of fundraising such as initial coin offerings (ICOs) and security token offerings (STOs), also sometimes referred to as digital security offerings (DSOs).[93] STO/DSOs may be conducted privately or on public, regulated stock exchange and are used to tokenize assets such as company shares, intellectual property, real estate, and others.[94]

    Games
    Main article: Blockchain game
    Blockchain technology, such as cryptocurrencies and non-fungible tokens (NFTs), has been used in video games for monetization. Many live-service games offer in-game customization options, such as character skins or other in-game items, which the players can earn and trade with other players using in-game currency. Some games also allow for trading of virtual items using real-world currency, but this may be illegal in some countries where video games are seen as akin to gambling, and has led to gray market issues such as skin gambling, and thus publishers typically have shied away from allowing players to earn real-world funds from games.[95] Blockchain games typically allow players to trade these in-game items for cryptocurrency, which can then be exchanged for money.[96]

    The first known game to use blockchain technologies was CryptoKitties, launched in November 2017, where the player would purchase NFTs with Ethereum cryptocurrency, each NFT consisting of a virtual pet that the player could breed with others to create offspring with combined traits as new NFTs.[97][96] The game made headlines in December 2017 when one virtual pet sold for more than US$100,000.[98] CryptoKitties also illustrated scalability problems for games on Ethereum when it created significant congestion on the Ethereum network in early 2018 with approximately 30% of all Ethereum transactions[clarification needed] being for the game.[99][100]

    By the early 2020s, there had not been a breakout success in video games using blockchain, as these games tend to focus on using blockchain for speculation instead of more traditional forms of gameplay, which offers limited appeal to most players. Such games also represent a high risk to investors as their revenues can be difficult to predict.[96] However, limited successes of some games, such as Axie Infinity during the COVID-19 pandemic, and corporate plans towards metaverse content, refueled interest in the area of GameFi, a term describing the intersection of video games and financing typically backed by blockchain currency, in the second half of 2021.[101] Several major publishers, including Ubisoft, Electronic Arts, and Take Two Interactive, have stated that blockchain and NFT-based games are under serious consideration for their companies in the future.[102]

    In October 2021, Valve Corporation banned blockchain games, including those using cryptocurrency and NFTs, from being hosted on its Steam digital storefront service, which is widely used for personal computer gaming, claiming that this was an extension of their policy banning games that offered in-game items with real-world value. Valve’s prior history with gambling, specifically skin gambling, was speculated to be a factor in the decision to ban blockchain games.[103] Journalists and players responded positively to Valve’s decision as blockchain and NFT games have a reputation for scams and fraud among most PC gamers,[95][103] and Epic Games, which runs the Epic Games Store in competition to Steam, said that they would be open to accepted blockchain games in the wake of Valve’s refusal.[104]

    Supply chain

    This section needs to be updated. Please help update this article to reflect recent events or newly available information. (August 2023)
    There have been several different efforts to employ blockchains in supply chain management.

    Precious commodities mining — Blockchain technology has been used for tracking the origins of gemstones and other precious commodities. In 2016, The Wall Street Journal reported that the blockchain technology company Everledger was partnering with IBM’s blockchain-based tracking service to trace the origin of diamonds to ensure that they were ethically mined.[105] As of 2019, the Diamond Trading Company (DTC) has been involved in building a diamond trading supply chain product called Tracer.[106]
    Food supply — As of 2018, Walmart and IBM were running a trial to use a blockchain-backed system for supply chain monitoring for lettuce and spinach – all nodes of the blockchain were administered by Walmart and located on the IBM cloud.[107]
    Fashion industry — There is an opaque relationship between brands, distributors, and customers in the fashion industry, which prevents the sustainable and stable development of the fashion industry. Blockchain could make this information transparent, assisting sustainable development of the industry.[108]
    Motor vehicles — Mercedes-Benz and partner Icertis developed a blockchain prototype used to facilitate consistent documentation of contracts along the supply chain so that the ethical standards and contractual obligations required of its direct suppliers can be passed on to second tier suppliers and beyond.[109][110] In another project, the company uses blockchain technology to track the emissions of climate-relevant gases and the amount of secondary material along the supply chain for its battery cell manufacturers.[111]
    Domain names
    There are several different efforts to offer domain name services via the blockchain. These domain names can be controlled by the use of a private key, which purports to allow for uncensorable websites. This would also bypass a registrar’s ability to suppress domains used for fraud, abuse, or illegal content.[112]

    Namecoin is a cryptocurrency that supports the “.bit” top-level domain (TLD). Namecoin was forked from bitcoin in 2011. The .bit TLD is not sanctioned by ICANN, instead requiring an alternative DNS root.[112] As of 2015, .bit was used by 28 websites, out of 120,000 registered names.[113] Namecoin was dropped by OpenNIC in 2019, due to malware and potential other legal issues.[114] Other blockchain alternatives to ICANN include The Handshake Network,[113] EmerDNS, and Unstoppable Domains.[112]

    Specific TLDs include “.eth”, “.luxe”, and “.kred”, which are associated with the Ethereum blockchain through the Ethereum Name Service (ENS). The .kred TLD also acts as an alternative to conventional cryptocurrency wallet addresses as a convenience for transferring cryptocurrency.[115]

    Other uses
    Blockchain technology can be used to create a permanent, public, transparent ledger system for compiling data on sales, tracking digital use and payments to content creators, such as wireless users[116] or musicians.[117] The Gartner 2019 CIO Survey reported 2% of higher education respondents had launched blockchain projects and another 18% were planning academic projects in the next 24 months.[118] In 2017, IBM partnered with ASCAP and PRS for Music to adopt blockchain technology in music distribution.[119] Imogen Heap’s Mycelia service has also been proposed as a blockchain-based alternative “that gives artists more control over how their songs and associated data circulate among fans and other musicians.”[120][121]

    New distribution methods are available for the insurance industry such as peer-to-peer insurance, parametric insurance and microinsurance following the adoption of blockchain.[122][123] The sharing economy and IoT are also set to benefit from blockchains because they involve many collaborating peers.[124] The use of blockchain in libraries is being studied with a grant from the U.S. Institute of Museum and Library Services.[125]

    Other blockchain designs include Hyperledger, a collaborative effort from the Linux Foundation to support blockchain-based distributed ledgers, with projects under this initiative including Hyperledger Burrow (by Monax) and Hyperledger Fabric (spearheaded by IBM).[126][127][128] Another is Quorum, a permissioned private blockchain by JPMorgan Chase with private storage, used for contract applications.[129]

    Oracle introduced a blockchain table feature in its Oracle 21c database.[68][69]

    Blockchain is also being used in peer-to-peer energy trading.[130][131][132]

    Blockchain and distributed ledger technologies have been investigated for managing access permissions, consent, provenance and audit trails in health care. A 2026 scoping review of 892 studies found that most involved prototypes or simulated evaluations, while evidence of sustained real-world implementation remained limited.[133]

    Lightweight blockchains, or simplified blockchains, are more suitable for internet of things (IoT) applications than conventional blockchains.[134] One experiment suggested that a lightweight blockchain-based network could accommodate up to 1.34 million authentication processes every second, which could be sufficient for resource-constrained IoT networks.[135]

    Blockchain could be used in detecting counterfeits by associating unique identifiers to products, documents and shipments, and storing records associated with transactions that cannot be forged or altered.[136][137] It is however argued that blockchain technology needs to be supplemented with technologies that provide a strong binding between physical objects and blockchain systems,[138] as well as provisions for content creator verification ala KYC standards.[139] The EUIPO established an Anti-Counterfeiting Blockathon Forum, with the objective of “defining, piloting and implementing” an anti-counterfeiting infrastructure at the European level.[140][141] The Dutch Standardisation organisation NEN uses blockchain together with QR Codes to authenticate certificates.[142]

    Beijing and Shanghai are among the cities designated by China to trial blockchain applications as January 30, 2022.[143] In Chinese legal proceedings, blockchain technology was first accepted as a method for authenticating internet evidence by the Hangzhou Internet Court in 2019 and has since been accepted by other Chinese courts.[144]: 123–125 

    Blockchain interoperability
    With the increasing number of blockchain systems appearing, even only those that support cryptocurrencies, blockchain interoperability is becoming a topic of major importance. The objective is to support transferring assets from one blockchain system to another blockchain system. Wegner[145] stated that “interoperability is the ability of two or more software components to cooperate despite differences in language, interface, and execution platform”. The objective of blockchain interoperability is therefore to support such cooperation among blockchain systems, despite those kinds of differences.

    There are already several blockchain interoperability solutions available.[146] They can be classified into three categories: cryptocurrency interoperability approaches, blockchain engines, and blockchain connectors.

    Several individual IETF participants produced the draft of a blockchain interoperability architecture.[147]

    Energy consumption concerns
    Some cryptocurrencies use blockchain mining, namely the peer-to-peer computer computations by which transactions are validated and verified. This requires a large amount of energy. In June 2018, the Bank for International Settlements criticized the use of public proof-of-work blockchains for their high energy consumption.[148][149][150]

    Early concern over the high energy consumption was a factor in later blockchains such as Cardano (2017), Solana (2020) and Polkadot (2020) adopting the less energy-intensive proof-of-stake model. Researchers have estimated that bitcoin consumes 100,000 times as much energy as proof-of-stake networks.[151][152]

    In 2021, a study by Cambridge University determined that bitcoin (at 121 terawatt-hours per year) used more electricity than Argentina (at 121TWh) and the Netherlands (109TWh).[153] According to Digiconomist, one bitcoin transaction required 708 kilowatt-hours of electrical energy, the amount an average U.S. household consumed in 24 days.[154]

    In February 2021, U.S. Treasury secretary Janet Yellen called bitcoin “an extremely inefficient way to conduct transactions”, saying “the amount of energy consumed in processing those transactions is staggering”.[155] In March 2021, Bill Gates stated that “Bitcoin uses more electricity per transaction than any other method known to mankind”, adding “It’s not a great climate thing.”[156]

    Nicholas Weaver, of the International Computer Science Institute at the University of California, Berkeley, examined blockchain’s online security, and the energy efficiency of proof-of-work public blockchains, and in both cases found it grossly inadequate.[157][158] The 31TWh-45TWh of electricity used for bitcoin in 2018 produced 17–23 million tonnes of CO2.[159][160] By 2022, the University of Cambridge and Digiconomist estimated that the two largest proof-of-work blockchains, bitcoin and Ethereum, together used twice as much electricity in one year as the whole of Sweden, leading to the release of up to 120 million tonnes of CO2 each year.[161]

    Some cryptocurrency developers are considering moving from the proof-of-work model to the proof-of-stake model.[162] In Sept, 2022, Ethereum converted from proof-of-work to proof-of-stake.[163]

    Academic research

    Blockchain panel discussion at the first IEEE Computer Society TechIgnite conference
    In October 2014, the MIT Bitcoin Club, with funding from MIT alumni, provided undergraduate students at the Massachusetts Institute of Technology access to $100 of bitcoin. The adoption rates, as studied by Catalini and Tucker (2016), revealed that when people who typically adopt technologies early are given delayed access, they tend to reject the technology.[164] Many universities have founded departments focusing on crypto and blockchain, including MIT, in 2017. In the same year, Edinburgh became “one of the first big European universities to launch a blockchain course”, according to the Financial Times.[165]

    Adoption decision
    Motivations for adopting blockchain technology (an aspect of innovation adoption) have been investigated by researchers. For example, Janssen, et al. provided a framework for analysis,[166] and Koens & Poll pointed out that adoption could be heavily driven by non-technical factors.[167] Based on behavioral models, Li[168] has discussed the differences between adoption at the individual level and organizational levels.

    Collaboration
    Scholars in business and management have started studying the role of blockchains to support collaboration.[169][170] It has been argued that blockchains can foster both cooperation (i.e., prevention of opportunistic behavior) and coordination (i.e., communication and information sharing). Thanks to reliability, transparency, traceability of records, and information immutability, blockchains facilitate collaboration in a way that differs both from the traditional use of contracts and from relational norms. Contrary to contracts, blockchains do not directly rely on the legal system to enforce agreements.[171] In addition, contrary to the use of relational norms, blockchains do not require a trust or direct connections between collaborators.

    Blockchain and internal audit
    External videos
    video icon Blockchain Basics & Cryptography, Gary Gensler, Massachusetts Institute of Technology, 0:30[172]
    The need for internal audits to provide effective oversight of organizational efficiency will require a change in the way that information is accessed in new formats.[173] Blockchain adoption requires a framework to identify the risk of exposure associated with transactions using blockchain. The Institute of Internal Auditors has identified the need for internal auditors to address this transformational technology. New methods are required to develop audit plans that identify threats and risks. The Internal Audit Foundation study, Blockchain and Internal Audit, assesses these factors.[174] The American Institute of Certified Public Accountants has outlined new roles for auditors as a result of blockchain.[175]

    Testnet
    In blockchain technology, a testnet is an instance of a blockchain powered by the same or a newer version of the underlying software, to be used for testing and experimentation without risk to real funds or the main chain.[176] Testnet coins are separate and distinct from the official mainnet coins, don’t have value, and can be obtained freely from faucets.[177]

    Testnets allow for the development of blockchain applications without the risk of losing funds.[178]

    Using testnets, a bug was discovered in the Bitcoin Core software that gave miners the ability to take down essential parts of the Bitcoin infrastructure (nodes) by sending a ‘bad’ block to the blockchain.[179]

    Mainnet
    A mainnet (short for main network) is the fully operational version of a blockchain where real transactions occur, as opposed to a testnet. It is secured through consensus mechanisms like Proof of Work or Proof of Stake and supports smart contracts, token transfers, and decentralized applications.[180]

    A mainnet launch marks the transition from a testnet to a live blockchain, involving security audits, network deployment, and token migration.[181]

    Journals
    Main article: Ledger (journal)
    In September 2015, the first peer-reviewed academic journal dedicated to cryptocurrency and blockchain technology research, Ledger, was announced. The inaugural issue was published in December 2016.[182] The journal covers aspects of mathematics, computer science, engineering, law, economics and philosophy that relate to cryptocurrencies.[183][184] The journal encourages authors to digitally sign a file hash of submitted papers, which are then timestamped into the bitcoin blockchain. Authors are also asked to include a personal bitcoin address on the first page of their papers for non-repudiation purposes.[

  • Computer security

    Computer security (also cybersecurity, digital security, or information technology (IT) security) is a subdiscipline within the field of information security. It focuses on protecting computer software, systems, and networks from threats that can lead to unauthorized information disclosure, theft, or damage to hardware, software, or data, as well as to the disruption or misdirection of the services they provide.[1][2]

    The growing significance of computer security reflects the increasing dependence on computer systems, the Internet,[3] and evolving wireless network standards. This reliance has expanded with the proliferation of smart devices, including smartphones, televisions, and other components of the Internet of things (IoT).

    As digital infrastructure becomes more embedded in everyday life, cybersecurity has emerged as a critical concern. The complexity of modern information systems—and the societal functions they underpin—has introduced new vulnerabilities. Systems that manage essential services, such as power grids, electoral processes, and finance, are particularly sensitive to security breaches.[4][5] Estimates of global information security spending in 2026 range from US$244 billion to US$308 billion.[6]

    Although many aspects of computer security involve digital security, such as electronic passwords and encryption, physical security measures, such as metal locks, are still used to prevent unauthorized tampering. IT security is not a perfect subset of information security and therefore does not completely align with the security convergence schema.

    Vulnerabilities and attacks
    Main article: Vulnerability (computing)
    A vulnerability refers to a flaw in the structure, execution, functioning, or internal oversight of a computer or system that compromises its security. Most of the vulnerabilities that have been discovered are documented in the Common Vulnerabilities and Exposures (CVE) database.[7] An exploitable vulnerability is one for which at least one working exploit exists.[8] A threat is an action or event that potentially affects hardware, software, or data negatively. A threat actor is a person or entity that performs a threat (the term hacker is more commonly used in popular media), likely by exploiting some vulnerability. Vulnerabilities can be researched, reverse-engineered, hunted, or exploited using automated tools or customized scripts.[9][10]

    Various people or parties are vulnerable to cyberattacks; however, different groups are likely to experience different types of attacks more than others.[11]

    In April 2023, the United Kingdom Department for Science, Innovation & Technology released a report on cyberattacks over the previous 12 months.[12] They surveyed 2,263 UK businesses, 1,174 UK registered charities, and 554 education institutions. The research found that “32% of businesses and 24% of charities overall recall any breaches or attacks from the last 12 months.” These figures were much higher for “medium businesses (59%), large businesses (69%), and high-income charities with £500,000 or more in annual income (56%).”[12] Yet, although medium or large businesses are more often the target, small and midsize businesses (SMBs) have become increasingly vulnerable as they often “do not have advanced tools to defend the business” .[11] SMBs are most likely to be affected by malware, ransomware, phishing, man-in-the-middle attacks, and Denial-of Service (DoS) Attacks.[11]

    Domestic internet users are most likely to be affected by untargeted cyberattacks.[13] These are where attackers indiscriminately target as many devices, services, or users as possible. They do this using techniques that take advantage of the openness of the Internet. These strategies mostly include phishing, ransomware, water holing and scanning.[13]

    To secure a computer system, it is important to understand the attacks that can be made against it, and these threats can typically be classified into one of the following categories:

    Backdoor
    A backdoor in a computer system, a cryptosystem or an algorithm, is any secret method of bypassing normal authentication or security controls. These weaknesses may exist for many reasons, including original design or poor configuration.[14] Due to the nature of backdoors, they are of greater concern to companies and databases as opposed to individuals.

    Backdoors may be added by an authorized party to allow some legitimate access or by an attacker for malicious reasons. Criminals often use malware to install backdoors, giving them remote administrative access to a system.[15] Once they have access, cybercriminals can “modify files, steal personal information, install unwanted software, and even take control of the entire computer.”[15]

    Backdoors can be difficult to detect, as they often remain hidden within source code or system firmware and may require intimate knowledge of the operating system to identify.

    Denial-of-service attack
    Denial-of-service attacks (DoS) are designed to make a machine or network resource unavailable to its intended users.[16] Attackers can deny service to individual victims, such as by deliberately entering an incorrect password enough consecutive times to cause the victim’s account to be locked, or they may overload the capabilities of a machine or network and block all users at once. While a network attack from a single IP address can be blocked by adding a new firewall rule, many forms of distributed denial-of-service (DDoS) attacks are possible, where the attack comes from a large number of points. In this case, defending against these attacks is much more difficult. Such attacks can originate from the zombie computers of a botnet or from a range of other possible techniques, including distributed reflective denial-of-service (DRDoS), where innocent systems are fooled into sending traffic to the victim.[16] With such attacks, the amplification factor makes the attack easier for the attacker because they have to use little bandwidth themselves. To understand why attackers may carry out these attacks, see the ‘attacker motivation’ section.

    Physical access attacks
    Main article: Physical access
    A direct-access attack is when an unauthorized user (an attacker) gains physical access to a computer, typically to copy data from it or steal information.[17] Attackers may also compromise security by making operating system modifications, installing software worms, keyloggers, covert listening devices or using wireless microphones. Even when the system is protected by standard security measures, these may be bypassed by booting another operating system or tool from a CD-ROM or other bootable media. Disk encryption and the Trusted Platform Module standard are designed to prevent these attacks.

    Direct service attackers are related in concept to direct memory attacks which allow an attacker to gain direct access to a computer’s memory.[18] The attacks “take advantage of a feature of modern computers that allows certain devices, such as external hard drives, graphics cards, or network cards, to access the computer’s memory directly.”[18]

    Eavesdropping
    Main article: Network eavesdropping
    Eavesdropping is the act of surreptitiously listening to a private computer conversation (communication), usually between hosts on a network. It typically occurs when a user connects to a network where traffic is not secured or encrypted and sends sensitive business data to a colleague, which, when listened to by an attacker, could be exploited.[19] Data transmitted across an open network can be intercepted by an attacker using various methods.

    Unlike malware, direct-access attacks, or other forms of cyberattacks, eavesdropping attacks are unlikely to negatively affect the performance of networks or devices, making them difficult to notice.[19] In fact, “the attacker does not need to have any ongoing connection to the software at all. The attacker can insert the software onto a compromised device, perhaps by direct insertion or perhaps by a virus or other malware, and then come back some time later to retrieve any data that is found or trigger the software to send the data at some determined time.”[20]

    Using a virtual private network (VPN), which encrypts data between two points, is one of the most common forms of protection against eavesdropping. Using the best form of encryption possible for wireless networks is best practice, as well as using HTTPS instead of an unencrypted HTTP.[21]

    Programs such as Carnivore and NarusInSight have been used by the Federal Bureau of Investigation (FBI) and the NSA to eavesdrop on the systems of internet service providers. Even machines that operate as a closed system (i.e., with no contact with the outside world) can be eavesdropped upon by monitoring the faint electromagnetic transmissions generated by the hardware. TEMPEST is a specification by the NSA referring to these attacks.

    Malware
    Main article: Malware
    Malicious software (malware) is any software code or computer program “intentionally written to harm a computer system or its users.”[22] Once present on a computer, it can leak sensitive details such as personal information, business information and passwords, can give control of the system to the attacker, and can corrupt or delete data permanently.[23][24]

    Types of malware
    Viruses are a specific type of malware, and are normally a malicious code that hijacks software with the intention to “do damage and spread copies of itself.” Copies are made with the aim of spreading to other programs on a computer.[22]
    Worms are similar to viruses, however viruses can only function when a user runs (opens) a compromised program. Worms are self-replicating malware that spread between programs, apps and devices without the need for human interaction.[22]
    Trojan horses are programs that pretend to be helpful or hide themselves within desired or legitimate software to “trick users into installing them.” Once installed, a RAT (Remote Access Trojan) can create a secret backdoor on the affected device to cause damage.[22]
    Spyware is a type of malware that secretly gathers information from an infected computer and transmits the sensitive information back to the attacker. One of the most common forms of spyware is keyloggers, which record all of a user’s keyboard inputs/keystrokes, to “allow hackers to harvest usernames, passwords, bank account and credit card numbers.”[22]
    Scareware, as the name suggests, is a form of malware that uses social engineering (manipulation) to scare, shock, trigger anxiety, or suggest the perception of a threat in order to manipulate users into buying or installing unwanted software. These attacks often begin with a “sudden pop-up with an urgent message, usually warning the user that they’ve broken the law or their device has a virus.”[22]
    Ransomware is when malware installs itself onto a victim’s machine, encrypts their files, and then turns around and demands a ransom (usually in Bitcoin) to return that data to the user.
    Man-in-the-middle attacks
    Man-in-the-middle attacks (MITM) involve a malicious attacker trying to intercept, surveil or modify communications between two parties by spoofing one or both party’s identities and injecting themselves in-between.[25] Types of MITM attacks include:

    IP address spoofing is where the attacker hijacks routing protocols to reroute the targets traffic to a vulnerable network node for traffic interception or injection.
    Message spoofing (via email, SMS or OTT messaging) is where the attacker spoofs the identity or carrier service while the target is using messaging protocols like email, SMS or OTT (IP-based) messaging apps. The attacker can then monitor conversations, launch social attacks or trigger zero-day-vulnerabilities to allow for further attacks.
    WiFi SSID spoofing is where the attacker simulates a Wi-Fi base station SSID to capture and modify internet traffic and transactions. The attacker can also use local network addressing and reduced network defenses to penetrate the target’s firewall by breaching known vulnerabilities. Sometimes known as a Pineapple attack thanks to a popular device. See also Malicious association.
    DNS spoofing is where attackers hijack domain name assignments to redirect traffic to systems under the attackers’ control, in order to surveil traffic or launch other attacks.
    SSL hijacking, typically coupled with another media-level MITM attack, is where the attacker spoofs the SSL authentication and encryption protocol by way of Certificate Authority injection in order to decrypt, surveil and modify traffic. See also TLS interception[25]
    Multi-vector, polymorphic attacks
    Surfacing in 2017, a new class of multi-vector,[26] polymorphic[27] cyber threats combine several types of attacks and change form to avoid cybersecurity controls as they spread.

    Multi-vector polymorphic attacks, as the name describes, are both multi-vectored and polymorphic.[28] Firstly, they are a singular attack that involves multiple methods of attack. In this sense, they are “multi-vectored” (i.e. the attack can use multiple means of propagation such as via the Web, email and applications). However, they are also multi-staged, meaning that “they can infiltrate networks and move laterally inside the network.”[28] The attacks can be polymorphic, meaning that the cyberattacks used such as viruses, worms or trojans “constantly change (“morph”) making it nearly impossible to detect them using signature-based defences.”[28]

    Phishing

    An example of a phishing email, disguised as an official email from a (fictional) bank. The sender is attempting to trick the recipient into revealing confidential information by confirming it at the phisher’s website. Note the misspelling of the words received and discrepancy as recieved and discrepency, respectively. Although the URL of the bank’s webpage appears to be legitimate, the hyperlink points at the phisher’s webpage.
    Phishing is the attempt to acquire sensitive information such as usernames, passwords, and credit card details directly from users by deceiving the users.[29] Phishing is typically carried out by email spoofing, instant messaging, text message, or on a phone call. They often direct users to enter details at a fake website whose look and feel are almost identical to the legitimate one.[30] The fake website often asks for personal information, such as login details and passwords. This information can then be used to gain access to the individual’s real account on the real website.

    Preying on a victim’s trust, phishing can be classified as a form of social engineering. Attackers can use creative ways to gain access to real accounts. A common scam is for attackers to send fake electronic invoices[31] to individuals showing that they recently purchased music, apps, or others, and instructing them to click on a link if the purchases were not authorized. A more strategic type of phishing is spear-phishing which leverages personal or organization-specific details to make the attacker appear like a trusted source. Spear-phishing attacks target specific individuals, rather than the broad net cast by phishing attempts.[32]

    Privilege escalation
    Privilege escalation describes a situation where an attacker with limited access is able, without authorization, to elevate their privileges or access level.[33] For example, a standard computer user may be able to exploit a vulnerability in the system to gain access to restricted data; or even become root and have full unrestricted access to a system. The severity of attacks can range from attacks simply sending an unsolicited email to a ransomware attack on large amounts of data. Privilege escalation usually starts with social engineering techniques, often phishing.[33]

    Privilege escalation can be separated into two strategies, horizontal and vertical privilege escalation:

    Horizontal escalation (or account takeover) is where an attacker gains access to a normal user account that has relatively low-level privileges. This may be through stealing the user’s username and password. Once they have access, they have gained a foothold, and using this foothold the attacker then may move around the network of users at this same lower level, gaining access to information of this similar privilege.[33]
    Vertical escalation, however, targets people higher up in a company and often with more administrative power, such as an employee in IT with a higher privilege. Using this privileged account will then enable the attacker to invade other accounts.[33]
    Side-channel attack
    Main article: Side-channel attack
    Any computational system affects its environment in some form. This effect it has on its environment can range from electromagnetic radiation, to residual effect on RAM cells which as a consequence make a cold boot attack possible, to hardware implementation faults that allow for access or guessing of other values that normally should be inaccessible. In Side-channel attack scenarios, the attacker would gather such information about a system or network to guess its internal state and as a result access the information which is assumed by the victim to be secure. The target information in a side channel can be challenging to detect due to its low amplitude when combined with other signals.[34]

    Social engineering
    Social engineering, in the context of computer security, aims to convince a user to disclose secrets such as passwords, card numbers, etc. or grant physical access by, for example, impersonating a senior executive, bank, a contractor, or a customer.[35] This generally involves exploiting people’s trust, and relying on their cognitive biases. A common scam involves emails sent to accounting and finance department personnel, impersonating their CEO and urgently requesting action. One of the main techniques of social engineering are phishing attacks.

    The term as it pertains to computer security was popularized by coverage of computer security consultant and convicted hacker Kevin Mitnick, who used social engineering to compromise secure information.[36] The prevalence of social engineering in the field of computer security has led to the development of the social cybersecurity discipline, which aims to study how social elements influence security changes and vice-versa.[37][38][39]

    In early 2016, the FBI reported that such business email compromise (BEC) scams had cost US businesses more than $2 billion in about two years.[40]

    In May 2016, the Milwaukee Bucks NBA team was the victim of this type of cyber scam with a perpetrator impersonating the team’s president Peter Feigin, resulting in the handover of all the team’s employees’ 2015 W-2 tax forms.[41]

    Spoofing
    Main article: Spoofing attack
    Spoofing is an act of pretending to be a valid entity through the falsification of data (such as an IP address or username), in order to gain access to information or resources that one is otherwise unauthorized to obtain. Spoofing is closely related to phishing.[42][43] There are several types of spoofing, including:

    Email spoofing is where an attacker forges the sending (From, or source) address of an email.
    IP address spoofing, where an attacker alters the source IP address in a network packet to hide their identity or impersonate another computing system.
    MAC spoofing, where an attacker modifies the Media Access Control (MAC) address of their network interface controller to obscure their identity, or to pose as another.
    Biometric spoofing, where an attacker produces a fake biometric sample to pose as another user.[44]
    Address Resolution Protocol (ARP) spoofing, where an attacker sends spoofed address resolution protocol onto a local area network to associate their Media Access Control address with a different host’s IP address. This causes data to be sent to the attacker rather than the intended host.
    Tampering
    Tampering describes a malicious modification or alteration of data. It is an intentional but unauthorized act resulting in the modification of a system, components of systems, its intended behavior, or data. So-called Evil Maid attacks and security services planting of surveillance capability into routers are examples.[45]

    HTML smuggling
    HTML smuggling allows an attacker to smuggle a malicious code inside a particular HTML or web page.[46] HTML files can carry payloads concealed as benign, inert data in order to defeat content filters. These payloads can be reconstructed on the other side of the filter.[47]

    When a target user opens the HTML, the malicious code is activated; the web browser then decodes the script, which then unleashes the malware onto the target’s device.[46]

    Information security practices
    Information security (InfoSec) and cybersecurity are closely related but not identical. While cybersecurity addresses external and malicious threats related to the exposure to the internet, information security also covers internal policies, roles, and controls. Employee behavior can have a big impact on information security in organizations. Cultural concepts can help different segments of the organization work effectively or work against effectiveness toward information security within an organization. Information security culture is the “…totality of patterns of behavior in an organization that contributes to the protection of information of all kinds.”[48]

    Andersson and Reimers (2014) found that employees often do not see themselves as part of their organization’s information security effort and often take actions that impede organizational changes.[49] Indeed, the Verizon Data Breach Investigations Report 2020, which examined 3,950 security breaches, discovered 30% of cybersecurity incidents involved internal actors within a company.[50] Research shows information security culture needs to be improved continuously. In “Information Security Culture from Analysis to Change”, authors commented, “It’s a never-ending process, a cycle of evaluation and change or maintenance.” To manage the information security culture, five steps should be taken: pre-evaluation, strategic planning, operative planning, implementation, and post-evaluation.[51]

    Pre-evaluation: To identify the awareness of information security within employees and to analyze the current security policies.
    Strategic planning: To develop an awareness program, clear targets need to be set. Assembling a team of skilled professionals is helpful to achieve it.
    Operative planning: Security culture can be established based on internal communication, management buy-in, security awareness and a training program.[51]
    Implementation: Four stages should be used to implement the information security culture. They are:
    Commitment of the management
    Communication with organizational members
    Courses for all organizational members
    Commitment of the employees[51]
    Post-evaluation: To assess the success of the planning and implementation, and to identify unresolved areas of concern.
    Computer protection (countermeasures)
    In computer security, a countermeasure is an action, device, procedure or technique that reduces a threat, a vulnerability, or an attack by eliminating or preventing it, by minimizing the harm it can cause, or by discovering and reporting it so that corrective action can be taken.[52][53][54]

    Some common countermeasures are listed in the following sections:

    Security by design
    Main article: Secure by design
    Security by design, or alternatively secure by design, means that the software has been designed from the ground up to be secure. In this case, security is considered a main feature.

    The UK government’s National Cyber Security Centre separates secure cyber design principles into five sections:[55]

    Before a secure system is created or updated, companies should ensure they understand the fundamentals and the context around the system they are trying to create and identify any weaknesses in the system.
    Companies should design and centre their security around techniques and defences which make attacking their data or systems inherently more challenging for attackers.
    Companies should ensure that their core services that rely on technology are protected so that the systems are essentially never down.
    Although systems can be created which are safe against a multitude of attacks, that does not mean that attacks will not be attempted. Despite one’s security, all companies’ systems should aim to be able to detect and spot attacks as soon as they occur to ensure the most effective response to them.
    Companies should create secure systems designed so that any attack that is successful has minimal severity.
    These design principles of security by design can include some of the following techniques:

    The principle of least privilege, where each part of the system has only the privileges that are needed for its function. That way, even if an attacker gains access to that part, they only have limited access to the whole system.
    Automated theorem proving to prove the correctness of crucial software subsystems.
    Code reviews and unit testing, approaches to make modules more secure where formal correctness proofs are not possible.
    Defense in depth, where the design is such that more than one subsystem needs to be violated to compromise the integrity of the system and the information it holds.
    Default secure settings, and design to fail secure rather than fail insecure (see fail-safe for the equivalent in safety engineering). Ideally, a secure system should require a deliberate, conscious, knowledgeable and free decision on the part of legitimate authorities in order to make it insecure.
    Audit trails track system activity so that when a security breach occurs, the mechanism and extent of the breach can be determined. Storing audit trails remotely, where they can only be appended to, can keep intruders from covering their tracks.
    Full disclosure of all vulnerabilities, to ensure that the window of vulnerability is kept as short as possible when bugs are discovered.
    Security architecture
    Security architecture can be defined as the “practice of designing computer systems to achieve security goals.”[56] These goals have overlap with the principles of “security by design” explored above, including to “make initial compromise of the system difficult,” and to “limit the impact of any compromise.”[56] In practice, the role of a security architect would be to ensure the structure of a system reinforces the security of the system, and that new changes are safe and meet the security requirements of the organization.[57][58]

    Similarly, Techopedia defines security architecture as “a unified security design that addresses the necessities and potential risks involved in a certain scenario or environment. It also specifies when and where to apply security controls. The design process is generally reproducible.” The key attributes of security architecture are:[59]

    the relationship of different components and how they depend on each other.
    determination of controls based on risk assessment, good practices, finances, and legal matters.
    the standardization of controls.
    Practicing security architecture provides the right foundation to systematically address business, IT and security concerns in an organization.

    Security measures
    A state of computer security is the conceptual ideal, attained by the use of three processes: threat prevention, detection, and response. These processes are based on various policies and system components, which include the following:

    Limiting the access of individuals using user account access controls and using cryptography can protect systems files and data, respectively.
    Firewalls are by far the most common prevention systems from a network security perspective as they can (if properly configured) shield access to internal network services and block certain kinds of attacks through packet filtering. Firewalls can be both hardware and software-based. Firewalls monitor and control incoming and outgoing traffic of a computer network and establish a barrier between a trusted network and an untrusted network.[60]
    Intrusion Detection System (IDS) products are designed to detect network attacks in-progress and assist in post-attack forensics, while audit trails and logs serve a similar function for individual systems.
    Response is necessarily defined by the assessed security requirements of an individual system and may cover the range from simple upgrade of protections to notification of legal authorities, counter-attacks, and the like. In some special cases, the complete destruction of the compromised system is favored, as it may happen that not all the compromised resources are detected.
    Cybersecurity awareness training to cope with cyber threats and attacks.[61]
    Forward web proxy solutions can prevent the client to visit malicious web pages and inspect the content before downloading to the client machines.
    Today, computer security consists mainly of preventive measures, like firewalls or an exit procedure. A firewall can be defined as a way of filtering network data between a host or a network and another network, such as the Internet. They can be implemented as software running on the machine, hooking into the network stack (or, in the case of most UNIX-based operating systems such as Linux, built into the operating system kernel) to provide real-time filtering and blocking.[60] Another implementation is a so-called physical firewall, which consists of a separate machine filtering network traffic. Firewalls are common amongst machines that are permanently connected to the Internet.

    Some organizations are turning to big data platforms, such as Apache Hadoop, to extend data accessibility and machine learning to detect advanced persistent threats.[62]

    In order to ensure adequate security, the confidentiality, integrity and availability of a network, known as the CIA triad, must be protected and is considered the foundation of information security.[63] To achieve those objectives, administrative, physical and technical security measures should be employed. The amount of security afforded to an asset can only be determined when its value is known.[64]

    Vulnerability management
    Main article: Vulnerability management
    Vulnerability management is the cycle of identifying, fixing or mitigating vulnerabilities,[65] especially in software and firmware. Vulnerability management is integral to computer security and network security.

    Vulnerabilities can be discovered with a vulnerability scanner, which analyzes a computer system in search of known vulnerabilities,[66] such as open ports, insecure software configuration, and susceptibility to malware. In order for these tools to be effective, they must be kept up to date with every new update the vendor releases. Typically, these updates will scan for the new vulnerabilities that were introduced recently.

    Beyond vulnerability scanning, many organizations contract outside security auditors to run regular penetration tests against their systems to identify vulnerabilities. In some sectors, this is a contractual requirement.[67]

    Reducing vulnerabilities
    The act of assessing and reducing vulnerabilities to cyberattacks is commonly referred to as information technology security assessments. They aim to assess systems for risk and to predict and test for their vulnerabilities. While formal verification of the correctness of computer systems is possible,[68][69] it is not yet common. Operating systems formally verified include seL4,[70] and SYSGO’s PikeOS[71][72] – but these make up a very small percentage of the market.

    It is possible to reduce an attacker’s chances by keeping systems up to date with security patches and updates and by hiring people with expertise in security. Large companies with significant threats can hire Security Operations Centre (SOC) Analysts. These are specialists in cyber defences, with their role ranging from “conducting threat analysis to investigating reports of any new issues and preparing and testing disaster recovery plans.”[73]

    Whilst no measures can completely guarantee the prevention of an attack, these measures can help mitigate the damage of possible attacks. The effects of data loss/damage can be also reduced by careful backing up and insurance.

    Outside of formal assessments, there are various methods of reducing vulnerabilities, including hardening systems.[74] Two factor authentication is a method for mitigating unauthorized access to a system or sensitive information.[75] It requires something you know: a password or PIN, and something you have: a card, dongle, cellphone, or another piece of hardware. This increases security as an unauthorized person needs both of these to gain access.

    Protecting against social engineering and direct computer access (physical) attacks can only happen by non-computer means, which can be difficult to enforce, relative to the sensitivity of the information. Training is often involved to help mitigate this risk by improving people’s knowledge of how to protect themselves and by increasing people’s awareness of threats.[76] However, even in highly disciplined environments (e.g., military organizations), social engineering attacks can still be difficult to foresee and prevent.

    Inoculation, derived from inoculation theory, seeks to prevent social engineering and other fraudulent tricks and traps by instilling a resistance to persuasion attempts through exposure to similar or related attempts.[77]

    Hardware protection mechanisms
    See also: Computer security compromised by hardware failure
    Hardware-based or assisted computer security also offers an alternative to software-only computer security. Using devices and methods such as dongles, trusted platform modules, intrusion-aware cases, drive locks, disabling USB ports, and mobile-enabled access may be considered more secure due to the physical access (or sophisticated backdoor access) required in order to be compromised. Each of these is covered in more detail below.

    USB dongles are typically used in software licensing schemes to unlock software capabilities,[78] but they can also be seen as a way to prevent unauthorized access to a computer or other device’s software. The dongle, or key, essentially creates a secure encrypted tunnel between the software application and the key. The principle is that an encryption scheme on the dongle, such as Advanced Encryption Standard (AES) provides a stronger measure of security since it is harder to hack and replicate the dongle than to simply copy the native software to another machine and use it. Another security application for dongles is to use them for accessing web-based content such as cloud software or Virtual Private Networks (VPNs).[79] In addition, a USB dongle can be configured to lock or unlock a computer.[80]
    Trusted platform modules (TPMs) secure devices by integrating cryptographic capabilities onto access devices, through the use of microprocessors, or so-called computers-on-a-chip. TPMs used in conjunction with server-side software offer a way to detect and authenticate hardware devices, preventing unauthorized network and data access.[81]
    Computer case intrusion detection refers to a device, typically a push-button switch, which detects when a computer case is opened. The firmware or BIOS is programmed to show an alert to the operator when the computer is booted up the next time.
    Drive locks are essentially software tools to encrypt hard drives, making them inaccessible to thieves.[82] Tools exist specifically for encrypting external drives as well.[83]
    Disabling USB ports is a security option for preventing unauthorized and malicious access to an otherwise secure computer. Infected USB dongles connected to a network from a computer inside the firewall are considered by the magazine Network World as the most common hardware threat facing computer networks.
    Disconnecting or disabling peripheral devices (like camera, GPS, removable storage, etc.), that are not in use.[84]
    Mobile-enabled access devices are growing in popularity due to the ubiquitous nature of cell phones.[85] Built-in capabilities such as Bluetooth, the newer Bluetooth low-energy (LE), near-field communication (NFC) on non-iOS devices and biometric validation such as thumbprint readers, as well as QR code reader software designed for mobile devices, offer new, secure ways for mobile phones to connect to access control systems. These control systems provide computer security and can also be used for controlling access to secure buildings.[86]
    IOMMUs allow for hardware-based sandboxing of components in mobile and desktop computers by utilizing direct memory access protections.[87][88]
    Physical Unclonable Functions (PUFs) can be used as a digital fingerprint or a unique identifier to integrated circuits and hardware, providing users the ability to secure the hardware supply chains going into their systems.[89][90]
    Secure operating systems
    Main article: Security-evaluated operating system
    One use of the term computer security refers to technology that is used to implement secure operating systems. Using secure operating systems is a good way of ensuring computer security. These are systems that have achieved certification from an external security-auditing organization, the most popular evaluations are Common Criteria (CC).[91]

    Secure coding
    Main article: Secure coding
    In software engineering, secure coding aims to guard against the accidental introduction of security vulnerabilities. It is also possible to create software designed from the ground up to be secure. Such systems are secure by design. Beyond this, formal verification aims to prove the correctness of the algorithms underlying a system;[92] important for cryptographic protocols for example.

    Capabilities and access control lists
    Main articles: Access control list, Role-based access control, and Capability-based security
    Within computer systems, two of the main security models capable of enforcing privilege separation are access control lists (ACLs) and role-based access control (RBAC).

    An access-control list (ACL), with respect to a computer file system, is a list of permissions associated with an object. An ACL specifies which users or system processes are granted access to objects, as well as what operations are allowed on given objects.

    Role-based access control is an approach to restricting system access to authorized users,[93][94][95] used by the majority of enterprises with more than 500 employees,[96] and can implement mandatory access control (MAC) or discretionary access control (DAC).

    A further approach, capability-based security has been mostly restricted to research operating systems. Capabilities can, however, also be implemented at the language level, leading to a style of programming that is essentially a refinement of standard object-oriented design. An open-source project in the area is the E language.

    User security training
    The end-user is widely recognized as the weakest link in the security chain[97] and it is estimated that more than 90% of security incidents and breaches involve some kind of human error.[98][99] Among the most commonly recorded forms of errors and misjudgment are poor password management, sending emails containing sensitive data and attachments to the wrong recipient, the inability to recognize misleading URLs and to identify fake websites and dangerous email attachments. A common mistake that users make is saving their user id/password in their browsers to make it easier to log in to banking sites. This is a gift to attackers who have obtained access to a machine by some means. The risk may be mitigated by the use of two-factor authentication.[100]

    As the human component of cyber risk is particularly relevant in determining the global cyber risk[101] an organization is facing, security awareness training, at all levels, not only provides formal compliance with regulatory and industry mandates but is considered essential[102] in reducing cyber risk and protecting individuals and companies from the great majority of cyber threats.

    The focus on the end-user represents a profound cultural change for many security practitioners, who have traditionally approached cybersecurity exclusively from a technical perspective, and moves along the lines suggested by major security centers[103] to develop a culture of cyber awareness within the organization, recognizing that a security-aware user provides an important line of defense against cyberattacks.

    Digital hygiene
    Related to end-user training, digital hygiene or cyber hygiene is a fundamental principle relating to information security and, as the analogy with personal hygiene shows, is the equivalent of establishing simple routine measures to minimize the risks from cyber threats. The assumption is that good cyber hygiene practices can give networked users another layer of protection, reducing the risk that one vulnerable node will be used to either mount attacks or compromise another node or network, especially from common cyberattacks.[104] Cyber hygiene should also not be mistaken for proactive cyber defence, a military term.[105]

    The most common acts of digital hygiene can include updating malware protection, cloud back-ups, passwords, and ensuring restricted admin rights and network firewalls.[12] As opposed to a purely technology-based defense against threats, cyber hygiene mostly regards routine measures that are technically simple to implement and mostly dependent on discipline[106] or education.[107] It can be thought of as an abstract list of tips or measures that have been demonstrated as having a positive effect on personal or collective digital security. As such, these measures can be performed by laypeople, not just security experts.

    Cyber hygiene relates to personal hygiene as computer viruses relate to biological viruses (or pathogens). However, while the term computer virus was coined almost simultaneously with the creation of the first working computer viruses,[108] the term cyber hygiene is a much later invention, perhaps as late as 2000[109] by Internet pioneer Vint Cerf. It has since been adopted by the Congress[110] and Senate of the United States,[111] the FBI,[112] EU institutions[104] and heads of state.[105]

    Difficulty of responding to breaches
    Responding to attempted security breaches is often very difficult for a variety of reasons, including:

    Identifying attackers is difficult, as they may operate through proxies, temporary anonymous dial-up accounts, wireless connections, and other anonymizing procedures which make back-tracing difficult – and are often located in another jurisdiction. If they successfully breach security, they have also often gained enough administrative access to enable them to delete logs to cover their tracks.
    The sheer number of attempted attacks, often by automated vulnerability scanners and computer worms, is so large that organizations cannot spend time pursuing each.
    Law enforcement officers often lack the skills, interest or budget to pursue attackers. Furthermore, identifying attackers across a network may necessitate collecting logs from multiple locations within the network and across various countries, a process that can be both difficult and time-consuming.
    Where an attack succeeds and a breach occurs, many jurisdictions now have in place mandatory security breach notification laws.

    Types of security and privacy
    Access control
    Anti-keyloggers
    Anti-malware
    Anti-spyware
    Anti-subversion software
    Anti-tamper software
    Anti-theft
    Antivirus software
    Cryptographic software
    Computer-aided dispatch (CAD)
    Data loss prevention software
    Firewall
    Intrusion detection system (IDS)
    Intrusion prevention system (IPS)
    Log management software
    Parental control
    Records management
    Sandbox
    Security information management
    Security information and event management (SIEM)
    Software and operating system updating
    Vulnerability management
    Systems at risk
    The growth in the number of computer systems and the increasing reliance upon them by individuals, businesses, industries, and governments means that there are an increasing number of systems at risk.

    Financial systems
    The computer systems of financial regulators and financial institutions like the U.S. Securities and Exchange Commission, SWIFT, investment banks, and commercial banks are prominent hacking targets for cybercriminals interested in manipulating markets and making illicit gains.[113] Websites and apps that accept or store credit card numbers, brokerage accounts, and bank account information are also prominent hacking targets, because of the potential for immediate financial gain from transferring money, making purchases, or selling the information on the black market.[114] In-store payment systems and ATMs have also been tampered with in order to gather customer account data and PINs.

    The UCLA Internet Report: Surveying the Digital Future (2000) found that the privacy of personal data created barriers to online sales and that more than nine out of 10 internet users were somewhat or very concerned about credit card security.[115]

    The most common web technologies for improving security between browsers and websites are named SSL (Secure Sockets Layer), and its successor TLS (Transport Layer Security), identity management and authentication services, and domain name services allow companies and consumers to engage in secure communications and commerce. Several versions of SSL and TLS are commonly used today in applications such as web browsing, e-mail, internet faxing, instant messaging, and VoIP (voice-over-IP). There are various interoperable implementations of these technologies, including at least one implementation that is open source. Open source allows anyone to view the application’s source code, and look for and report vulnerabilities.

    The credit card companies Visa and MasterCard cooperated to develop the secure EMV chip which is embedded in credit cards. Further developments include the Chip Authentication Program where banks give customers hand-held card readers to perform online secure transactions. Other developments in this arena include the development of technology such as Instant Issuance which has enabled shopping mall kiosks acting on behalf of banks to issue on-the-spot credit cards to interested customers.

    Utilities and industrial equipment
    Computers control functions at many utilities, including coordination of telecommunications, the power grid, nuclear power plants, and valve opening and closing in water and gas networks. The Internet is a potential attack vector for such machines if connected, but the Stuxnet worm demonstrated that even equipment controlled by computers not connected to the Internet can be vulnerable. In 2014, the Computer Emergency Readiness Team, a division of the Department of Homeland Security, investigated 79 hacking incidents at energy companies.[116]

    Aviation
    The aviation industry is very reliant on a series of complex systems which could be attacked.[117] A simple power outage at one airport can cause repercussions worldwide,[118] much of the system relies on radio transmissions which could be disrupted,[119] and controlling aircraft over oceans is especially dangerous because radar surveillance only extends 175 to 225 miles offshore.[120] There is also potential for attack from within an aircraft.[121]

    Implementing fixes in aerospace systems poses a unique challenge because efficient air transportation is heavily affected by weight and volume. Improving security by adding physical devices to airplanes could increase their unloaded weight, and could potentially reduce cargo or passenger capacity.[122]

    In Europe, with the (Pan-European Network Service)[123] and NewPENS,[124] and in the US with the NextGen program,[125] air navigation service providers are moving to create their own dedicated networks.

    Many modern passports are now biometric passports, containing an embedded microchip that stores a digitized photograph and personal information such as name, gender, and date of birth. In addition, more countries[which?] are introducing facial recognition technology to reduce identity-related fraud. The introduction of the ePassport has assisted border officials in verifying the identity of the passport holder, thus allowing for quick passenger processing.[126] Plans are under way in the US, the UK, and Australia to introduce SmartGate kiosks with both retina and fingerprint recognition technology.[127] The airline industry is moving from the use of traditional paper tickets towards the use of electronic tickets (e-tickets). These have been made possible by advances in online credit card transactions in partnership with the airlines. Long-distance bus companies[which?] are also switching over to e-ticketing transactions today.

    The consequences of a successful attack range from loss of confidentiality to loss of system integrity, air traffic control outages, loss of aircraft, and even loss of life.

    Consumer devices
    Desktop computers and laptops are commonly targeted to gather passwords or financial account information or to construct a botnet to attack another target. Smartphones, tablet computers, smart watches, and other mobile devices such as quantified self devices like activity trackers have sensors such as cameras, microphones, GPS receivers, compasses, and accelerometers which could be exploited, and may collect personal information, including sensitive health information. WiFi, Bluetooth, and cell phone networks on any of these devices could be used as attack vectors, and sensors might be remotely activated after a successful breach.[128]

    The increasing number of home automation devices such as the Nest thermostat are also potential targets.[128]

    Healthcare
    Today many healthcare providers and health insurance companies use the internet to provide enhanced products and services. Examples are the use of tele-health to potentially offer better quality and access to healthcare, or fitness trackers to lower insurance premiums.[129] Patient records are increasingly being placed on secure in-house networks, alleviating the need for extra storage space.[130]

    Large corporations
    Large corporations are common targets. In many cases attacks are aimed at financial gain through identity theft and involve data breaches. Examples include the loss of millions of clients’ credit card and financial details by Home Depot,[131] Staples,[132] Target Corporation,[133] and Equifax.[134]

    Medical records have been targeted in general identify theft, health insurance fraud, and impersonating patients to obtain prescription drugs for recreational purposes or resale.[135] Although cyber threats continue to increase, 62% of all organizations did not increase security training for their business in 2015.[136]

    Not all attacks are financially motivated, however: security firm HBGary Federal had a serious series of attacks in 2011 from hacktivist group Anonymous in retaliation for the firm’s CEO claiming to have infiltrated their group,[137][138] and Sony Pictures was hacked in 2014 with the apparent dual motive of embarrassing the company through data leaks and crippling the company by wiping workstations and servers.[139][140]

    Automobiles
    See also: Autonomous car § Potential disadvantages, Automated driving system § Risks and liabilities, and Automotive hacking
    Vehicles are increasingly computerized, with engine timing, cruise control, anti-lock brakes, seat belt tensioners, door locks, airbags and advanced driver-assistance systems on many models. Additionally, connected cars may use WiFi and Bluetooth to communicate with onboard consumer devices and the cell phone network.[141] Self-driving cars are expected to be even more complex. All of these systems carry some security risks, and such issues have gained wide attention.[142][143][144]

    Simple examples of risk include a malicious compact disc being used as an attack vector,[145] and the car’s onboard microphones being used for eavesdropping. However, if access is gained to a car’s internal controller area network, the danger is much greater[141] – and in a widely publicized 2015 test, hackers remotely carjacked a vehicle from 10 miles away and drove it into a ditch.[146][147]

    Manufacturers are reacting in numerous ways, with Tesla in 2016 pushing out some security fixes over the air into its cars’ computer systems.[148] In the area of autonomous vehicles, in September 2016 the United States Department of Transportation announced some initial safety standards, and called for states to come up with uniform policies.[149][150][151]

    Additionally, e-Drivers’ licenses are being developed using the same technology. For example, Mexico’s licensing authority (ICV) has used a smart card platform to issue the first e-Drivers’ licenses to the city of Monterrey, in the state of Nuevo León.[152]

    Shipping
    Shipping companies[153] have adopted RFID (Radio Frequency Identification) technology as an efficient, digitally secure, tracking device. Unlike a barcode, RFID can be read up to 20 feet away. RFID is used by FedEx[154] and UPS.[155]

    Government
    Government and military computer systems are commonly attacked by activists[156][157][158] and foreign powers.[159][160][161][162] This includes local and regional government infrastructure such as traffic light controls, police and intelligence agency communications, personnel records, as well as student records.[163]

    Internet of things and physical vulnerabilities
    The Internet of things (IoT) is the network of physical objects such as devices, vehicles, and buildings that are embedded with electronics, software, sensors, and network connectivity that enables them to collect and exchange data.[164] Concerns have been raised that this is being developed without appropriate consideration of the security challenges involved.[165][166]

    While the IoT creates opportunities for more direct integration of the physical world into computer-based systems,[167][168] it also provides opportunities for misuse. In particular, as the Internet of Things spreads widely, cyberattacks are likely to become an increasingly physical (rather than simply virtual) threat.[169] If a front door’s lock is connected to the Internet, and can be locked/unlocked from a phone, then a criminal could enter the home at the press of a button from a stolen or hacked phone. People could stand to lose much more than their credit card numbers in a world controlled by IoT-enabled devices. Thieves have also used electronic means to circumvent non-Internet-connected hotel door locks.[170]

    An attack aimed at physical infrastructure or human lives is often called a cyber-kinetic attack. As IoT devices and appliances become more widespread, the prevalence and potential damage of cyber-kinetic attacks can increase substantially.

    Medical systems
    See also: Medical device hijack and Medical data breach
    Medical devices have either been successfully attacked or had potentially deadly vulnerabilities demonstrated, including both in-hospital diagnostic equipment[171] and implanted devices including pacemakers[172] and insulin pumps.[173] There are many reports of hospitals and hospital organizations getting hacked, including ransomware attacks,[174][175][176][177] Windows XP exploits,[178][179] viruses,[180][181] and data breaches of sensitive data stored on hospital servers.[182][175][183][184] On 28 December 2016 the US Food and Drug Administration released its recommendations for how medical device manufacturers should maintain the security of Internet-connected devices – but no structure for enforcement.[185][186]

    Energy sector
    In distributed generation systems, the risk of a cyberattack is real, according to Daily Energy Insider. An attack could cause a loss of power in a large area for a long period of time, and such an attack could have just as severe consequences as a natural disaster. The District of Columbia is considering creating a Distributed Energy Resources (DER) Authority within the city, with the goal being for customers to have more insight into their own energy use and giving the local electric utility, Pepco, the chance to better estimate energy demand. The D.C. proposal, however, would “allow third-party vendors to create numerous points of energy distribution, which could potentially create more opportunities for cyberattackers to threaten the electric grid.”[187]

    Telecommunications
    Perhaps the most widely known digitally secure telecommunication device is the SIM (Subscriber Identity Module) card, a device that is embedded in most of the world’s cellular devices before any service can be obtained. The SIM card is just the beginning of this digitally secure environment.

    The Smart Card Web Servers draft standard (SCWS) defines the interfaces to an HTTP server in a smart card.[188] Tests are being conducted to secure OTA (“over-the-air”) payment and credit card information from and to a mobile phone. Combination SIM/DVD devices are being developed through Smart Video Card technology which embeds a DVD-compliant optical disc into the card body of a regular SIM card.

    Other telecommunication developments involving digital security include mobile signatures, which use the embedded SIM card to generate a legally binding electronic signature.

    Cost and impact of security breaches
    Serious financial damage has been caused by security breaches, but because there is no standard model for estimating the cost of an incident, the only data available is that which is made public by the organizations involved. “Several computer security consulting firms produce estimates of total worldwide losses attributable to virus and worm attacks and to hostile digital acts in general. The 2003 loss estimates by these firms range from $13 billion (worms and viruses only) to $226 billion (for all forms of covert attacks). The reliability of these estimates is often challenged; the underlying methodology is basically anecdotal.”[189]

    However, reasonable estimates of the financial cost of security breaches can actually help organizations make rational investment decisions. According to the classic Gordon-Loeb Model analyzing the optimal investment level in information security, one can conclude that the amount a firm spends to protect information should generally be only a small fraction of the expected loss (i.e., the expected value of the loss resulting from a cyber/information security breach).[190]

    Attacker motivation
    As with physical security, the motivations for breaches of computer security vary between attackers. Some are thrill-seekers or vandals, some are activists, others are criminals looking for financial gain. State-sponsored attackers are now common and well resourced but started with amateurs such as Markus Hess who hacked for the KGB, as recounted by Clifford Stoll in The Cuckoo’s Egg.

    Attackers motivations can vary for all types of attacks from pleasure to political goals.[16] For example, hacktivists may target a company or organization that carries out activities they do not agree with. This would be to create bad publicity for the company by having its website crash.

    High capability hackers, often with larger backing or state sponsorship, may attack based on the demands of their financial backers. These attacks are more likely to attempt more serious attack. An example of a more serious attack was the 2015 Ukraine power grid hack, which reportedly utilised the spear-phishing, destruction of files, and denial-of-service attacks to carry out the full attack.[191][192]

    Additionally, recent attacker motivations can be traced back to extremist organizations seeking to gain political advantage or disrupt social agendas.[193] The growth of the internet, mobile technologies, and inexpensive computing devices have led to a rise in capabilities but also to the risk to environments that are deemed as vital to operations. All critical targeted environments are susceptible to compromise and this has led to a series of proactive studies on how to migrate the risk by taking into consideration motivations by these types of actors. Several stark differences exist between the hacker motivation and that of nation state actors seeking to attack based on an ideological preference.[194]

    A key aspect of threat modeling for any system is identifying the motivations behind potential attacks and the individuals or groups likely to carry them out. The level and detail of security measures will differ based on the specific system being protected. For instance, a home personal computer, a bank, and a classified military network each face distinct threats, despite using similar underlying technologies.[195]

    Computer security incident management
    Computer security incident management is an organized approach to addressing and managing the aftermath of a computer security incident or compromise with the goal of preventing a breach or thwarting a cyberattack. An incident that is not identified and managed at the time of intrusion typically escalates to a more damaging event such as a data breach or system failure. The intended outcome of a computer security incident response plan is to contain the incident, limit damage and assist recovery to business as usual. Responding to compromises quickly can mitigate exploited vulnerabilities, restore services and processes and minimize losses.[196] Incident response planning allows an organization to establish a series of best practices to stop an intrusion before it causes damage. Typical incident response plans contain a set of written instructions that outline the organization’s response to a cyberattack. Without a documented plan in place, an organization may not successfully detect an intrusion or compromise and stakeholders may not understand their roles, processes and procedures during an escalation, slowing the organization’s response and resolution.

    There are four key components of a computer security incident response plan:

    Preparation: Preparing stakeholders on the procedures for handling computer security incidents or compromises
    Detection and analysis: Identifying and investigating suspicious activity to confirm a security incident, prioritizing the response based on impact and coordinating notification of the incident
    Containment, eradication and recovery: Isolating affected systems to prevent escalation and limit impact, pinpointing the genesis of the incident, removing malware, affected systems and bad actors from the environment and restoring systems and data when a threat no longer remains
    Post incident activity: Post mortem analysis of the incident, its root cause and the organization’s response with the intent of improving the incident response plan and future response efforts.[197]
    Notable attacks and breaches
    Further information: List of cyberattacks and List of data breaches
    Some illustrative examples of different types of computer security breaches are given below.

    Robert Morris and the first computer worm
    Main article: Morris worm
    In 1988, 60,000 computers were connected to the Internet, and most were mainframes, minicomputers and professional workstations. On 2 November 1988, many started to slow down, because they were running a malicious code that demanded processor time and that spread itself to other computers – the first internet computer worm.[198] The software was traced back to 23-year-old Cornell University graduate student Robert Tappan Morris who said “he wanted to count how many machines were connected to the Internet”.[198]

    Rome Laboratory
    In 1994, over a hundred intrusions were made by unidentified crackers into the Rome Laboratory, the US Air Force’s main command and research facility. Using trojan horses, hackers were able to obtain unrestricted access to Rome’s networking systems and remove traces of their activities. The intruders were able to obtain classified files, such as air tasking order systems data and furthermore able to penetrate connected networks of National Aeronautics and Space Administration’s Goddard Space Flight Center, Wright-Patterson Air Force Base, some Defense contractors, and other private sector organizations, by posing as a trusted Rome center user.[199]

    TJX customer credit card details
    In early 2007, American apparel and home goods company TJX announced that it was the victim of an unauthorized computer systems intrusion[200] and that the hackers had accessed a system that stored data on credit card, debit card, check, and merchandise return transactions.[201]

    Stuxnet attack
    In 2010, the computer worm known as Stuxnet reportedly ruined almost one-fifth of Iran’s nuclear centrifuges.[202] It did so by disrupting industrial programmable logic controllers (PLCs) in a targeted attack. This is generally believed to have been launched by Israel and the United States to disrupt Iran’s nuclear program[203][204][205][206] – although neither has publicly admitted this.

    Global surveillance disclosures
    Main article: Global surveillance disclosures (2013–present)
    In early 2013, documents provided by Edward Snowden were published by The Washington Post and The Guardian[207][208] exposing the massive scale of NSA global surveillance. There were also indications that the NSA may have inserted a backdoor in a NIST standard for encryption.[209] This standard was later withdrawn due to widespread criticism.[210] The NSA additionally were revealed to have tapped the links between Google’s data centers.[211]

    Target and Home Depot breaches
    A Ukrainian hacker known as Rescator broke into Target Corporation computers in 2013, stealing roughly 40 million credit cards,[212] and then Home Depot computers in 2014, stealing between 53 and 56 million credit card numbers.[213] Warnings were delivered at both corporations, but ignored; physical security breaches using self checkout machines are believed to have played a large role. “The malware utilized is absolutely unsophisticated and uninteresting,” says Jim Walter, director of threat intelligence operations at security technology company McAfee – meaning that the heists could have easily been stopped by existing antivirus software had administrators responded to the warnings. The size of the thefts has resulted in major attention from state and Federal United States authorities and the investigation is ongoing.

    Office of Personnel Management data breach
    In April 2015, the Office of Personnel Management discovered it had been hacked more than a year earlier in a data breach, resulting in the theft of approximately 21.5 million personnel records handled by the office.[214] The Office of Personnel Management hack has been described by federal officials as among the largest breaches of government data in the history of the United States.[215] Data targeted in the breach included personally identifiable information such as Social Security numbers, names, dates and places of birth, addresses, and fingerprints of current and former government employees as well as anyone who had undergone a government background check.[216][217] It is believed the hack was perpetrated by Chinese hackers.[218]

    Ashley Madison breach
    Main article: Ashley Madison Data Breach
    In July 2015, a hacker group known as The Impact Team successfully breached the extramarital relationship website Ashley Madison, created by Avid Life Media. The group claimed that they had taken not only company data but user data as well. After the breach, The Impact Team dumped emails from the company’s CEO, to prove their point, and threatened to dump customer data unless the website was taken down permanently.[219] When Avid Life Media did not take the site offline the group released two more compressed files, one 9.7GB and the second 20GB. After the second data dump, Avid Life Media CEO Noel Biderman resigned; but the website remained to function.

    Colonial Pipeline ransomware attack
    Main article: Colonial Pipeline ransomware attack
    In June 2021, the cyberattack took down the largest fuel pipeline in the U.S. and led to shortages across the East Coast.[220]

    Legal issues and global regulation
    International legal issues of cyberattacks are complicated in nature. There is no global base of common rules to judge, and eventually punish, cybercrimes and cybercriminals – and where security firms or agencies do locate the cybercriminal behind the creation of a particular piece of malware or form of cyberattack, often the local authorities cannot take action due to lack of laws under which to prosecute.[221][222] Proving attribution for cybercrimes and cyberattacks is also a major problem for all law enforcement agencies. “Computer viruses switch from one country to another, from one jurisdiction to another – moving around the world, using the fact that we don’t have the capability to globally police operations like this. So the Internet is as if someone [had] given free plane tickets to all the online criminals of the world.”[221] The use of techniques such as dynamic DNS, fast flux and bullet proof servers add to the difficulty of investigation and enforcement.

    Role of government
    The role of the government is to make regulations to force companies and organizations to protect their systems, infrastructure and information from any cyberattacks, but also to protect its own national infrastructure such as the national power-grid.[223]

    The government’s regulatory role in cyberspace is complicated. For some, cyberspace was seen as a virtual space that was to remain free of government intervention, as can be seen in many of today’s libertarian blockchain and bitcoin discussions.[224]

    Many government officials and experts think that the government should do more and that there is a crucial need for improved regulation, mainly due to the failure of the private sector to solve efficiently the cybersecurity problem. R. Clarke said during a panel discussion at the RSA Security Conference in San Francisco, he believes that the “industry only responds when you threaten regulation. If the industry doesn’t respond (to the threat), you have to follow through.”[225] On the other hand, executives from the private sector agree that improvements are necessary, but think that government intervention would affect their ability to innovate efficiently. Daniel R. McCarthy analyzed this public-private partnership in cybersecurity and reflected on the role of cybersecurity in the broader constitution of political order.[226]

    On 22 May 2020, the UN Security Council held its second ever informal meeting on cybersecurity to focus on cyber challenges to international peace. According to UN Secretary-General António Guterres, new technologies are too often used to violate rights.[227]

    International actions
    Many different teams and organizations exist, including:

    The Forum of Incident Response and Security Teams (FIRST) is the global association of CSIRTs.[228] The US-CERT, AT&T, Apple, Cisco, McAfee, Microsoft are all members of this international team.[229]
    The Council of Europe helps protect societies worldwide from the threat of cybercrime through the Convention on Cybercrime.[230]
    The purpose of the Messaging Anti-Abuse Working Group (MAAWG) is to bring the messaging industry together to work collaboratively and to successfully address the various forms of messaging abuse, such as spam, viruses, denial-of-service attacks and other messaging exploitations.[231] France Telecom, Facebook, AT&T, Apple, Cisco, Sprint are some of the members of the MAAWG.[232]
    ENISA: The European Union Agency for Cybersecurity (ENISA), established in 2004 as the European Network and Information Security Agency, is an agency of the European Union whose mandate under Regulation (EU) 2019/881 (the Cybersecurity Act) is to help achieve “a high common level of cybersecurity across the Union”.[233] Under the NIS2 Directive it provides the secretariat of the network of national computer security incident response teams (CSIRTs network) and of EU-CyCLONe and maintains a European vulnerability database, while supervision of the entities covered by the directive rests with national competent authorities.[234][235]
    Europe
    On 14 April 2016, the European Parliament and the Council of the European Union adopted the General Data Protection Regulation (GDPR). The GDPR, which came into force on 25 May 2018, grants individuals within the European Union (EU) and the European Economic Area (EEA) the right to the protection of personal data. The regulation requires that any entity that processes personal data incorporate data protection by design and by default. It also requires that certain organizations appoint a Data Protection Officer (DPO).

    The IT Security Association TeleTrusT exist in Germany since June 1986, which is an international competence network for IT security.

    National actions
    To protect national network and system security, many countries have established strategies and staffing for computer emergency response teams, proactive cyber defence, and cyber threat intelligence. National policy actions typically include cybersecurity regulations and information security standards.

    Canada
    Since 2010, Canada has had a cyber security strategy.[236][237] This functions as a counterpart document to the National Strategy and Action Plan for Critical Infrastructure.[238] The strategy has three main pillars: securing government systems, securing vital private cyber systems, and helping Canadians to be secure online.[237][238] There is also a Cyber Incident Management Framework to provide a coordinated response in the event of a cyber incident.[239][238]

    The Canadian Cyber Incident Response Centre (CCIRC) is responsible for mitigating and responding to threats to Canada’s critical infrastructure and cyber systems. It provides support to mitigate cyber threats, technical support to respond & recover from targeted cyber attacks, and provides online tools for members of Canada’s critical infrastructure sectors.[240] It posts regular cyber security bulletins[241] & operates an online reporting tool where individuals and organizations can report a cyber incident.[242]

    To inform the general public on how to protect themselves online, Public Safety Canada has partnered with STOP.THINK.CONNECT, a coalition of non-profit, private sector, and government organizations,[243] and launched the Cyber Security Cooperation Program.[244] They also run the GetCyberSafe portal for Canadian citizens, and Cyber Security Awareness Month during October.[245]

    Public Safety Canada aims to begin an evaluation of Canada’s cybersecurity strategy in early 2015.[238]

    Australia
    Australian federal government announced an $18.2 million investment to fortify the cyber security resilience of small and medium enterprises (SMEs) and enhance their capabilities in responding to cyber threats. This financial backing is an integral component of the 2023-2030 Australian Cyber Security Strategy.[246] A substantial allocation of $7.2 million is earmarked for the establishment of a voluntary cyber health check program, facilitating businesses in conducting a comprehensive and tailored self-assessment of their cyber security upskill.

    This avant-garde health assessment serves as a diagnostic tool, enabling enterprises to ascertain the robustness of Australia’s cyber security regulations.[247] Furthermore, it affords them access to a repository of educational resources and materials, fostering the acquisition of skills necessary for an elevated cyber security posture. This groundbreaking initiative was jointly disclosed by Minister for Cyber Security Clare O’Neil and Minister for Small Business Julie Collins.[248]

    Hong Kong
    Hong Kong’s Protection of Critical Infrastructures (Computer Systems) Bill (the “Bill”) was passed by the Legislative Council on 19 March 2025, with the purpose to “establish legal requirements for organisations designated as critical infrastructure operators”.[249] To defend the economy and public safety against the cyber threats of severe disruption, Hong Kong’s new Protection of Critical Infrastructures (Computer Systems) Ordinance (Cap.653) (Ordinance), together with its Code of Practice (CoP) guidelines for gatekeepers at the front line of defence, came into effect on 1 January 2026.[250]

    India
    Some provisions for cybersecurity have been incorporated into rules framed under the Information Technology Act 2000.[251]

    The National Cyber Security Policy 2013 is a policy framework by the Ministry of Electronics and Information Technology (MeitY) which aims to protect the public and private infrastructure from cyberattacks, and safeguard “information, such as personal information (of web users), financial and banking information and sovereign data”. CERT- In is the nodal agency which monitors the cyber threats in the country. The post of National Cyber Security Coordinator has also been created in the Prime Minister’s Office (PMO).

    The Indian Companies Act 2013 has also introduced cyber law and cyber security obligations on the part of Indian directors. Some provisions for cyber security have been incorporated into rules framed under the Information Technology Act 2000 Update in 2013.[252]

    South Korea
    Following cyberattacks in the first half of 2013, when the government, news media, television stations, and bank websites were compromised, the national government committed to the training of 5,000 new cybersecurity experts by 2017. The South Korean government blamed its northern counterpart for these attacks, as well as incidents that occurred in 2009, 2011,[253] and 2012, but Pyongyang denies the accusations.[254]

    United Kingdom
    In 2016 the National Cyber Security Centre was formed as the central body overseeing cybersecurity in the UK, as part of GCHQ.[255][256] The UK government published a National Cyber Security Strategy in 2022 assigning £2.6bn for industry, skills and national security.[257][258] In addition, the National Cyber Force, launched in 2020, works with GCHQ and the Ministry of Defence and aims to “transform the UK’s ability to contest adversaries in cyber space, to protect the country, its people and our way of life”.[259]

    United States
    Strategies and directives
    In 2013, executive order 13636 Improving Critical Infrastructure Cybersecurity was signed, which prompted the creation of the NIST Cybersecurity Framework.

    The 2018 cyber strategy called for specific measures to harden U.S. government networks from attacks, such as the June 2015 intrusion into the U.S. Office of Personnel Management (OPM), which compromised the records of about 4.2 million current and former government employees.[260]

    In response to the Colonial Pipeline ransomware attack President Joe Biden signed Executive Order 14028 on May 12, 2021, to increase software security standards for sales to the government, tighten detection and security on existing systems, improve information sharing and training, establish a Cyber Safety Review Board, and improve incident response.[261][262]

    The Biden administration released a comprehensive National Cybersecurity Strategy in 2023.[263]

    Legislation
    The 1986 Computer Fraud and Abuse Act prohibits unauthorized access or damage of protected computers as defined in 18 U.S.C. § 1030(e)(2).

    Standardized government testing services
    The General Services Administration (GSA) has[when?] standardized the penetration test service as a pre-vetted support service, to rapidly address potential vulnerabilities, and stop adversaries before they impact US federal, state and local governments. These services are commonly referred to as Highly Adaptive Cybersecurity Services (HACS).

    Further information: Penetration test § Standardized government penetration test services
    Agencies
    The Department of Homeland Security has a dedicated division responsible for the response system, risk management program and requirements for cyber security in the United States called the National Cyber Security Division.[264][265] The division is home to US-CERT operations and the National Cyber Alert System.[265] The National Cybersecurity and Communications Integration Center brings together government organizations responsible for protecting computer networks and networked infrastructure.[266]

    The third priority of the FBI is to: “Protect the United States against cyber-based attacks and high-technology crimes”,[267] and they, along with the National White Collar Crime Center (NW3C), and the Bureau of Justice Assistance (BJA) are part of the multi-agency task force, The Internet Crime Complaint Center, also known as IC3.[268]

    In addition to its own specific duties, the FBI participates alongside non-profit organizations such as InfraGard.[269][270]

    The Computer Crime and Intellectual Property Section (CCIPS) operates in the United States Department of Justice Criminal Division. The CCIPS is in charge of investigating computer crime and intellectual property crime and is specialized in the search and seizure of digital evidence in computers and networks.[271] In 2017, CCIPS published A Framework for a Vulnerability Disclosure Program for Online Systems to help organizations “clearly describe authorized vulnerability disclosure and discovery conduct, thereby substantially reducing the likelihood that such described activities will result in a civil or criminal violation of law under the Computer Fraud and Abuse Act (18 U.S.C. § 1030).”[272]

    The United States Cyber Command, also known as USCYBERCOM, “has the mission to direct, synchronize, and coordinate cyberspace planning and operations to defend and advance national interests in collaboration with domestic and international partners.”[273] It has no role in the protection of civilian networks.[274][275]

    The U.S. Federal Communications Commission’s role in cybersecurity is to strengthen the protection of critical communications infrastructure, to assist in maintaining the reliability of networks during disasters, to aid in swift recovery after, and to ensure that first responders have access to effective communications services.[276]

    The Food and Drug Administration has issued guidance for medical devices,[277] and the National Highway Traffic Safety Administration[278] is concerned with automotive cyber security. After being criticized by the Government Accountability Office,[279] and following successful attacks on airports and claimed attacks on airplanes, the Federal Aviation Administration has devoted funding to securing systems on board the planes of private manufacturers, and the Aircraft Communications Addressing and Reporting System.[280] Concerns have also been raised about the future Next Generation Air Transportation System.[281]

    The US Department of Defense (DoD) issued DoD Directive 8570 in 2004, supplemented by DoD Directive 8140, requiring all DoD employees and all DoD contract personnel involved in information assurance roles and activities to earn and maintain various industry Information Technology (IT) certifications in an effort to ensure that all DoD personnel involved in network infrastructure defense have minimum levels of IT industry recognized knowledge, skills and abilities (KSA). Andersson and Reimers (2019) report these certifications range from CompTIA’s A+ and Security+ through the ICS2.org’s CISSP, etc.[282]

    Computer emergency response team is a name given to expert groups that handle computer security incidents. In the US, two distinct organizations exist, although they do work closely together.

    US-CERT: part of the National Cyber Security Division of the United States Department of Homeland Security.[283]
    CERT/CC: created by the Defense Advanced Research Projects Agency (DARPA) and run by the Software Engineering Institute (SEI).
    In the context of U.S. nuclear power plants, the U.S. Nuclear Regulatory Commission (NRC) outlines cyber security requirements under 10 CFR Part 73, specifically in 10 CFR 73.54. The Nuclear Energy Institute’s NEI 08-09 document, Cyber Security Plan for Nuclear Power Reactors, outlines a comprehensive framework for cybersecurity in the nuclear power industry.[284]

    Modern warfare
    Main article: Cyberwarfare
    There is growing concern that cyberspace will become the next theater of warfare. As Mark Clayton from The Christian Science Monitor wrote in a 2015 article titled “The New Cyber Arms Race”:

    In the future, wars will not just be fought by soldiers with guns or with planes that drop bombs. They will also be fought with the click of a mouse a half a world away that unleashes carefully weaponized computer programs that disrupt or destroy critical industries like utilities, transportation, communications, and energy. Such attacks could also disable military networks that control the movement of troops, the path of jet fighters, the command and control of warships.[285]

    This has led to new terms such as cyberwarfare and cyberterrorism. The United States Cyber Command was created in 2009[286] and many other countries have similar forces.

    There are a few critics that question whether cyber security is as significant a threat as it is made out to be.[287][288][289]

    Careers
    Cyber security is a fast-growing field of IT concerned with reducing organizations’ risk of getting hacked or data breaches.[290] According to research from the Enterprise Strategy Group, 46% of organizations say that they have a “problematic shortage” of cyber security skills in 2016, up from 28% in 2015.[291] Commercial, government and non-governmental organizations all employ cybersecurity professionals. The fastest increases in demand for cyber security workers are in industries managing increasing volumes of consumer data such as finance, health care, and retail.[292] However, the use of the term cybersecurity is more prevalent in government job descriptions.[293]

    Cybersecurity job titles and descriptions include:[294]

    Security analyst
    Analyzes and assesses vulnerabilities in the infrastructure (software, hardware, networks), investigates using available tools and countermeasures to remedy the detected vulnerabilities and recommends solutions and best practices. Analyzes and assesses damage to the data/infrastructure as a result of security incidents, examines available recovery tools and processes, and recommends solutions. Tests for compliance with security policies and procedures. May assist in the creation, implementation, or management of security solutions.
    Security engineer
    Performs security monitoring, security and data/logs analysis, and forensic analysis, to detect security incidents, and mount the incident response. Investigates and utilizes new technologies and processes to enhance security capabilities and implement improvements. May also review code or perform other security engineering methodologies.
    Security architect
    Designs a security system or major components of a security system, and may head a security design team building a new security system.[295]
    Chief Information Security Officer (CISO)
    A high-level management position responsible for the entire information security division/staff. The position may include hands-on technical work.[296]
    Chief Security Officer (CSO)
    A high-level management position responsible for the entire security division/staff. A newer position is now deemed needed as security risks grow.
    Data Protection Officer (DPO)
    A DPO is tasked with monitoring compliance with data protection laws (such as GDPR), data protection policies, awareness-raising, training, and audits.[297]
    Security consultant/specialist/intelligence
    Broad titles that encompass any one or all of the other roles or titles tasked with protecting computers, networks, software, data or information systems against viruses, worms, spyware, malware, intrusion detection, unauthorized access, denial-of-service attacks, and an ever-increasing list of attacks by hackers acting as individuals or as part of organized crime or foreign governments.
    Student programs are also available for people interested in beginning a career in cyber security.[298][299] Meanwhile, an alternative option for information security professionals of varied experience levels to keep studying is online security training, including webcasts.[300][301] A wide range of courses are also available.[302]

    In the United Kingdom, a nationwide set of cyber security forums, known as the U.K Cyber Security Forum, were established supported by the Government’s cyber security strategy[303] in order to encourage start-ups and innovation and to address the skills gap[304] identified by the U.K Government.

    In Singapore, the Cyber Security Agency has issued a Singapore Operational Technology (OT) Cybersecurity Competency Framework (OTCCF). The framework defines emerging cybersecurity roles in Operational Technology. The OTCCF was endorsed by the Infocomm Media Development Authority (IMDA). It outlines the different OT cybersecurity job positions as well as the technical skills and core competencies necessary. It also depicts the many career paths available, including vertical and lateral advancement opportunities.[305]

    Terminology
    The following terms used with regards to computer security are explained below:

    Access authorization restricts access to a computer to a group of users through the use of authentication systems. These systems can protect either the whole computer, such as through an interactive login screen, or individual services, such as a FTP server. There are many methods for identifying and authenticating users, such as passwords, identification cards, smart cards, and biometric systems.
    Anti-virus software consists of computer programs that attempt to identify, thwart, and eliminate computer viruses and other malicious software (malware).
    Applications are executable code, so general corporate practice is to restrict or block users the power to install them; to install them only when there is a demonstrated need (e.g. software needed to perform assignments); to install only those which are known to be reputable (preferably with access to the computer code used to create the application), and to reduce the attack surface by installing as few as possible. They are typically run with least privilege, with a robust process in place to identify, test and install any released security patches or updates for them.
    For example, programs can be installed into an individual user’s account, which limits the program’s potential access, as well as being a means control which users have specific exceptions to policy. In Linux, FreeBSD, OpenBSD, and other Unix-like operating systems there is an option to further restrict an application using chroot or other means of restricting the application to its own ‘sandbox’. For example. Linux provides namespaces, and Cgroups to further restrict the access of an application to system resources.
    Generalized security frameworks such as SELinux or AppArmor help administrators control access.
    Java and other languages which compile to Java byte code and run in the Java virtual machine can have their access to other applications controlled at the virtual machine level.
    Some software can be run in software containers which can even provide their own set of system libraries, limiting the software’s, or anyone controlling it, access to the server’s versions of the libraries.
    Authentication techniques can be used to ensure that communication end-points are who they say they are.
    Automated theorem proving and other verification tools can be used to enable critical algorithms and code used in secure systems to be mathematically proven to meet their specifications.
    Backups are one or more copies kept of important computer files. Typically, multiple copies will be kept at different locations so that if a copy is stolen or damaged, other copies will still exist.
    Capability and access control list techniques can be used to ensure privilege separation and mandatory access control. Capabilities vs. ACLs discusses their use.
    Chain of trust techniques can be used to attempt to ensure that all software loaded has been certified as authentic by the system’s designers.
    Confidentiality is the nondisclosure of information except to another authorized person.[306]
    Cryptographic techniques can be used to defend data in transit between systems, reducing the probability that the data exchange between systems can be intercepted or modified.
    Cyber attribution, is an attribution of cybercrime, i.e., finding who perpetrated a cyberattack.
    Cyberwarfare is an Internet-based conflict that involves politically motivated attacks on information and information systems. Such attacks can, for example, disable official websites and networks, disrupt or disable essential services, steal or alter classified data, and cripple financial systems.
    Data integrity is the accuracy and consistency of stored data, indicated by an absence of any alteration in data between two updates of a data record.[307]

    Cryptographic techniques involve transforming information, scrambling it, so it becomes unreadable during transmission. The intended recipient can unscramble the message; ideally, eavesdroppers cannot.
    Encryption is used to protect the confidentiality of a message. Cryptographically secure ciphers are designed to make any practical attempt of breaking them infeasible. Symmetric-key ciphers are suitable for bulk encryption using shared keys, and public-key encryption using digital certificates can provide a practical solution for the problem of securely communicating when no key is shared in advance.
    Endpoint security software aids networks in preventing malware infection and data theft at network entry points made vulnerable by the prevalence of potentially infected devices such as laptops, mobile devices, and USB drives.[308]
    Firewalls serve as a gatekeeper system between networks, allowing only traffic that matches defined rules. They often include detailed logging, and may include intrusion detection and intrusion prevention features. They are near-universal between company local area networks and the Internet, but can also be used internally to impose traffic rules between networks if network segmentation is configured.
    A hacker is someone who seeks to breach defenses and exploit weaknesses in a computer system or network.
    Honey pots are computers that are intentionally left vulnerable to attack by crackers. They can be used to catch crackers and to identify their techniques.
    Intrusion-detection systems are devices or software applications that monitor networks or systems for malicious activity or policy violations.
    A microkernel is an approach to operating system design which has only the near-minimum amount of code running at the most privileged level – and runs other elements of the operating system such as device drivers, protocol stacks and file systems, in the safer, less privileged user space.
    Pinging. The standard ping application can be used to test if an IP address is in use. If it is, attackers may then try a port scan to detect which services are exposed.
    A port scan is used to probe an IP address for open ports to identify accessible network services and applications.
    A key logger is spyware that silently captures and stores each keystroke that a user types on the computer’s keyboard.
    Social engineering is the use of deception to manipulate individuals to breach security.
    Logic bombs is a type of malware added to a legitimate program that lies dormant until it is triggered by a specific event.
    A unikernel is a computer program that runs on a minimalistic operating system where a single application is allowed to run (as opposed to a general purpose operating system where many applications can run at the same time). This approach to minimizing the attack surface is adopted mostly in cloud environments where software is deployed in virtual machines.
    Zero trust security means that no one is trusted by default from inside or outside the network, and verification is required from everyone trying to gain access to resources on the network.
    History
    Since the Internet’s arrival and with the digital transformation initiated in recent years, the notion of cybersecurity has become a familiar subject in both our professional and personal lives. Cybersecurity and cyber threats have been consistently present for the last 60 years of technological change. In the 1970s and 1980s, computer security was mainly limited to academia until the conception of the Internet, where, with increased connectivity, computer viruses and network intrusions began to take off. After the spread of viruses in the 1990s, the 2000s marked the institutionalization of organized attacks such as distributed denial of service.[309] This led to the formalization of cybersecurity as a professional discipline.[310]

    The April 1967 session organized by Willis Ware at the Spring Joint Computer Conference, and the later publication of the Ware Report, were foundational moments in the history of the field of computer security.[311] Ware’s work straddled the intersection of material, cultural, political, and social concerns.[311]

    A 1977 NIST publication[312] introduced the CIA triad of confidentiality, integrity, and availability as a clear and simple way to describe key security goals.[313] While still relevant, many more elaborate frameworks have since been proposed.[314][315]

    However, in the 1970s and 1980s, there were no grave computer threats because computers and the internet were still in the early stages of development, and security threats were easily identifiable. More often, threats came from malicious insiders who gained unauthorized access to sensitive documents and files. Although malware and network breaches existed during the early years, they did not use them for financial gain. By the second half of the 1970s, established computer firms like IBM started offering commercial access control systems and computer security software products.[316]

    One of the earliest examples of an attack on a computer network was the computer worm Creeper written by Bob Thomas at BBN, which propagated through the ARPANET in 1971.[317] The program was purely experimental in nature and carried no malicious payload. A later program, Reaper, was created by Ray Tomlinson in 1972 and used to destroy Creeper.[318]

    Between September 1986 and June 1987, a group of German hackers performed the first documented case of cyber espionage.[319] The group hacked into American defense contractors, universities, and military base networks and sold gathered information to the Soviet KGB. The group was led by Markus Hess, who was arrested on 29 June 1987. He was convicted of espionage (along with two co-conspirators) on 15 Feb 1990.

    In 1988, one of the first computer worms, called the Morris worm, was distributed via the Internet. It gained significant mainstream media attention.[320]

    Netscape started developing the protocol SSL, shortly after the National Center for Supercomputing Applications (NCSA) launched Mosaic 1.0, the first web browser, in 1993.[321][322] Netscape had SSL version 1.0 ready in 1994, but it was never released to the public due to many serious security vulnerabilities.[321] However, in 1995, Netscape launched Version 2.0.[323]

    The National Security Agency (NSA) is responsible for the protection of U.S. information systems and also for collecting foreign intelligence.[324] The agency analyzes commonly used software and system configurations to find security flaws, which it can use for offensive purposes against competitors of the United States.[325]

    NSA contractors created and sold click-and-shoot attack tools to US agencies and close allies, but eventually, the tools made their way to foreign adversaries.[326] In 2016, NSA’s own hacking tools were hacked, and Russia and North Korea have used it.[327] NSA’s employees and contractors have been recruited at high salaries by adversaries, anxious to compete in cyberwarfare.[328] In 2007, the United States and Israel began exploiting security flaws in the Microsoft Windows operating system to attack and damage equipment used in Iran to refine nuclear materials. Iran responded by heavily investing in their own cyberwarfare capability, which it began using against the United States.[325]

  • Machine learning

    For the journal, see Machine Learning (journal).
    “Statistical learning” redirects here. For statistical learning in linguistics, see Statistical learning in language acquisition.
    Part of a series on
    Machine learning
    and
    data mining
    Paradigms
    Problems
    Supervised learning
    (classification • regression)
    Clustering
    Dimensionality reduction
    Structured prediction
    Anomaly detection
    Neural networks
    Reinforcement learning
    Learning with humans
    Model diagnostics
    Mathematical foundations
    Journals and conferences
    Related articles
    vte
    Part of a series on
    Artificial intelligence (AI)
    Major goals
    Approaches
    Machine learningData miningSymbolicDeep learningShallow learningBayesian networksEvolutionary algorithmsNeuro-symbolic AISystems integrationOpen-sourceOpen weightsAI data centers
    Applications
    Philosophy
    History
    Controversies
    Security Incidents
    Glossary
    vte
    Machine learning (ML) is a field of study in artificial intelligence concerned with the development and study of statistical algorithms that can learn from data and generalize to unseen data, and thus perform tasks without being explicitly programmed.[1]

    Statistics and mathematical optimisation methods compose the foundations of machine learning. Data mining is a related field of study, focusing on exploratory data analysis (EDA) through unsupervised learning.[3][4]

    From a theoretical viewpoint, probably approximately correct learning provides a mathematical and statistical framework for describing machine learning. Most traditional machine learning and deep learning algorithms can be described as empirical risk minimisation under this framework.

    Advances in the field of deep learning have allowed neural networks, a class of statistical algorithms, to surpass many previous machine learning approaches in performance. The resulting dominance of machine learning within the field of AI from the late 2000s and early 2010s onwards was such that the terms “machine learning” and “artificial intelligence” are now often used interchangeably, even though they are not synonymous.[5]

    History
    For a chronological guide, see Timeline of machine learning.
    The term machine learning was coined in 1959 by Arthur Samuel, an IBM employee and pioneer in the field of computer gaming and artificial intelligence.[6][7] The synonym self-teaching computers was also used during this time period.[8][9]

    The earliest machine learning program was introduced in the 1950s, when Samuel invented a computer program that calculated the chance of winning in checkers for each side, but the history of machine learning is rooted in decades of efforts to study human cognitive processes.[10] In 1949, Canadian psychologist Donald Hebb published the book The Organization of Behavior, in which he introduced a theoretical neural structure formed by certain interactions among nerve cells.[11] The Hebbian theory of neuron interaction set the groundwork for how many machine learning algorithms work, with connected artificial neurons changing the strength of their connections based on data.[10] Other researchers who have studied human cognitive systems contributed to the modern machine learning technologies as well, including Walter Pitts and Warren McCulloch, who proposed the first mathematical model of neural networks including algorithms that mirror human thought processes.[10][verification needed]

    By the early 1960s, an experimental “learning machine” with punched tape memory, called Cybertron, had been developed by Raytheon Company to analyse sonar signals, electrocardiograms, and speech patterns using rudimentary reinforcement learning. It was repetitively “trained” by a human operator/teacher to recognise patterns and equipped with a “goof” button to cause it to reevaluate incorrect decisions.[12][importance?] A representative book[citation needed] on research into machine learning during the 1960s was Nils Nilsson’s book “Learning Machines”, dealing mostly with machine learning for pattern classification.[13] Interest related to pattern recognition continued into the 1970s, as described by Duda and Hart in 1973.[14] In 1981, a report was given on using teaching strategies so that an artificial neural network learns to recognise 40 characters (26 letters, 10 digits, and 4 special symbols) from a computer terminal.[15]

    Tom M. Mitchell provided a widely quoted,[citation needed] more formal definition of the algorithms studied in the machine learning field: “A computer program is said to learn from experience E with respect to some class of tasks T and performance measure P if its performance at tasks in T, as measured by P, improves with experience E.”[16] This definition of the tasks in which machine learning is concerned is fundamentally operational rather than defining the field in cognitive terms. This follows Alan Turing’s proposal in his paper “Computing Machinery and Intelligence”, in which the question, “Can machines think?”, is replaced by asking whether machines can convincingly imitate a human in its responses to human-posed questions.[17][18]

    In 2012, AlexNet, developed by Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton, achieved substantially improved results in the ImageNet image recognition competition, contributing to the wider adoption of deep neural networks.[19] In 2013, Tomáš Mikolov and colleagues introduced word2vec, techniques for efficiently learning distributed vector representations of words from large text corpora.[20][21] In 2014, Ian Goodfellow and colleagues introduced generative adversarial networks (GANs), a framework for training generative models through an adversarial process.[22] In 2016, AlphaGo became the first computer program to defeat a professional human Go player without handicaps on a full-sized board, using deep neural networks and reinforcement learning.[23] In 2017, Ashish Vaswani and colleagues introduced the Transformer, a neural network architecture based primarily on attention rather than recurrence or convolution.[24]

    Relationships to other fields
    Artificial intelligence

    Deep learning is a subset of machine learning, which is itself a subset of artificial intelligence.[25]
    As a scientific endeavour, machine learning grew out of the quest for artificial intelligence (AI). In the early days of AI as an academic discipline, some researchers were interested in having machines learn from data. They attempted to approach the problem with various symbolic methods, as well as what were then termed “neural networks”; these were mostly perceptrons and other models that were later found to be reinventions of the generalised linear models of statistics.[26] Probabilistic reasoning was also employed, especially in automated medical diagnosis.[27]: 488 

    However, an increasing emphasis on the logical, knowledge-based approach caused a rift between AI and machine learning. Probabilistic systems were plagued by theoretical and practical problems of data acquisition and representation.[27]: 488  By 1980, expert systems had come to dominate AI, and statistics was out of favour.[28] Work on symbolic/knowledge-based learning continued within AI, leading to inductive logic programming (ILP), but the more statistical line of research was now outside the field of AI proper, in pattern recognition and information retrieval.[27]: 708–710, 755  Neural network research was abandoned by AI and computer science around the same time. This subfield, termed “connectionism”, was continued by researchers from other disciplines, including John Hopfield, David Rumelhart, and Geoffrey Hinton. Their main success came in the mid-1980s with the reinvention of backpropagation.[27]: 25 

    Machine learning (ML), reorganised and recognised as its own field, started to flourish in the 1990s. The field changed its goal from achieving artificial intelligence to tackling solvable problems of a practical nature. It shifted focus away from the symbolic approaches it had inherited from AI, and toward methods and models borrowed from statistics, fuzzy logic, and probability theory.[28]

    Data compression
    This section is an excerpt from Data compression § Machine learning.edit
    There is a close connection between machine learning and compression. A system that predicts the posterior probabilities of a sequence given its entire history can be used for optimal data compression (by using arithmetic coding on the output distribution). Conversely, an optimal compressor can be used for prediction (by finding the symbol that compresses best, given the previous history). This equivalence has been used as a justification for using data compression as a benchmark for “general intelligence”.[29][30][31]

    An alternative view can show compression algorithms implicitly map strings into implicit feature space vectors, and compression-based similarity measures compute similarity within these feature spaces. For each compressor C(.) we define an associated vector space ℵ, such that C(.) maps an input string x, corresponding to the vector norm ||~x||. An exhaustive examination of the feature spaces underlying all compression algorithms is precluded by space; instead, feature vectors choose to examine three representative lossless compression methods, LZW, LZ77, and PPM.[32]

    According to AIXI theory, a connection more directly explained in Hutter Prize, the best possible compression of x is the smallest possible software that generates x. For example, in that model, a zip file’s compressed size includes both the zip file and the unzipping software, since you can not unzip it without both, but there may be an even smaller combined form.

    Examples of AI-powered audio/video compression software include NVIDIA Maxine, AIVC.[33] Examples of software that can perform AI-powered image compression include OpenCV, TensorFlow, MATLAB’s Image Processing Toolbox (IPT) and High-Fidelity Generative Image Compression.[34]

    In unsupervised machine learning, k-means clustering can be utilized to compress data by grouping similar data points into clusters. This technique simplifies handling extensive datasets that lack predefined labels and finds widespread use in fields such as image compression.[35]

    Data compression aims to reduce the size of data files, enhancing storage efficiency and speeding up data transmission. K-means clustering, an unsupervised machine learning algorithm, is employed to partition a dataset into a specified number of clusters, k, each represented by the centroid of its points. This process condenses extensive datasets into a more compact set of representative points. Particularly beneficial in image and signal processing, k-means clustering aids in data reduction by replacing groups of data points with their centroids, thereby preserving the core information of the original data while significantly decreasing the required storage space.[36]

    Large language models (LLMs) are also efficient lossless data compressors on some data sets, as demonstrated by DeepMind’s research with the Chinchilla 70B model. Developed by DeepMind, Chinchilla 70B effectively compressed data, outperforming conventional methods such as Portable Network Graphics (PNG) for images and Free Lossless Audio Codec (FLAC) for audio. It achieved compression of image and audio data to 43.4% and 16.4% of their original sizes, respectively. There is, however, some reason to be concerned that the data set used for testing overlaps the LLM training data set, making it possible that the Chinchilla 70B model is only an efficient compression tool on data it has already been trained on.[37][38]

    Data mining
    Machine learning and data mining often employ the same methods and overlap significantly, but while machine learning focuses on prediction based on known properties learned from the training data, data mining focuses on the discovery of previously unknown properties in the data (this is the analysis step of knowledge discovery in databases). Data mining uses many machine learning methods, but with different goals; on the other hand, machine learning also employs data mining methods as “unsupervised learning” or as a preprocessing step to improve learner accuracy. Much of the confusion between these two research communities comes from the basic assumptions they work with: in machine learning, performance is usually evaluated with respect to the ability to reproduce known knowledge, while in knowledge discovery and data mining (KDD) the key task is the discovery of previously unknown knowledge. Evaluated with respect to known knowledge, an uninformed (unsupervised) method will easily be outperformed by other supervised methods, while in a typical KDD task, supervised methods cannot be used due to the unavailability of training data.[citation needed]

    Machine learning also has intimate ties to optimization: Many learning problems are formulated as minimisation of some loss function on a training set of examples. Loss functions express the discrepancy between the predictions of the model being trained and the actual problem instances (for example, in classification, one wants to assign a label to instances, and models are trained to correctly predict the preassigned labels of a set of examples).[39]

    Generalization
    Characterizing the generalisation of various learning algorithms is an active topic of current research, especially for deep learning algorithms.[citation needed]

    Statistics
    Machine learning and statistics are closely related fields in terms of methods, but distinct in their principal goal: statistics draws population inferences from a sample, while machine learning finds generalisable predictive patterns.[40]

    Conventional statistical analyses require the a priori selection of a model most suitable for the study data set. In addition, only significant or theoretically relevant variables based on previous experience are included for analysis. In contrast, machine learning is not built on a pre-structured model; rather, the data shape the model by detecting underlying patterns. The more variables (input) used to train the model, the more accurate the ultimate model will be.[41]

    Leo Breiman distinguished two statistical modelling paradigms: the data model and the algorithmic model,[42] wherein “algorithmic model” means more or less the machine learning algorithms like Random forest.[clarification needed]

    Some statisticians have adopted methods from machine learning, producing the field of statistical learning.[43]

    Statistical physics
    Analytical and computational techniques derived from deep-rooted physics of disordered systems can be extended to large-scale problems, including machine learning, e.g., to analyse the weight space of deep neural networks.[44] Statistical physics is thus finding applications in the area of medical diagnostics.[45][clarification needed]

    Theory
    Main articles: Computational learning theory and Statistical learning theory
    A core objective of a learner is to generalise from its experience.[2][46] Generalisation in this context is the ability of a learning machine to perform accurately on new, unseen examples/tasks after having experienced a learning data set. The training examples come from some generally unknown probability distribution (considered representative of the space of occurrences) and the learner has to build a general model about this space that enables it to produce sufficiently accurate predictions in new cases.

    The computational analysis of machine learning algorithms and their performance is a branch of theoretical computer science known as computational learning theory. One major framework is the probably approximately correct learning model. Because training sets are finite and the future is uncertain, learning theory usually does not yield guarantees of the performance of algorithms. Instead, probabilistic bounds on the performance are quite common. The bias–variance decomposition is one way to quantify generalisation error.[citation needed]

    For the best performance in the context of generalisation, the complexity of the hypothesis should match the complexity of the function underlying the data. If the hypothesis is less complex than the function, then the model has underfitted the data. If the complexity of the model is increased in response, then the training error decreases. But if the hypothesis is too complex, then the model is subject to overfitting and generalisation will be poorer.[47]

    In addition to performance bounds, learning theorists study the time complexity and feasibility of learning. In computational learning theory, a computation is considered feasible if it can be done in polynomial time. There are two kinds of time complexity results: Positive results show that a certain class of functions can be learned in polynomial time. Negative results show that certain classes cannot be learned in polynomial time.[citation needed]

    Approaches

    In supervised learning, the training data is labelled with the expected answers, while in unsupervised learning, the model identifies patterns or structures in unlabelled data.
    Machine learning approaches are traditionally divided into three broad categories, which correspond to learning paradigms, depending on the nature of the “signal” or “feedback” available to the learning system:

    Supervised learning: The computer is presented with example inputs and their desired outputs, given by a “teacher”, and the goal is to learn a general rule that maps inputs to outputs.
    Unsupervised learning: No labels are given to the learning algorithm, leaving it on its own to find structure in its input. Unsupervised learning can be a goal in itself (discovering hidden patterns in data) or a means towards an end (feature learning).
    Reinforcement learning: A computer program interacts with a dynamic environment in which it must perform a certain goal (such as driving a vehicle or playing a game against an opponent). As it navigates its problem space, the program is provided rewards as feedback, which it tries to maximize, thus resulting in the program learning from experience.[2]
    Although each algorithm has advantages and limitations, no single algorithm works for all problems.[48][49][50]

    Supervised learning
    Main article: Supervised learning

    A support-vector machine is a supervised learning model that divides the data into regions separated by a linear boundary. Here, the linear boundary divides the black circles from the white.
    Supervised learning algorithms build a mathematical model of a set of data that contains both the inputs and the desired outputs.[51] The data, known as training data, consists of a set of training examples. Each training example has one or more inputs and the desired output, also known as a supervisory signal. In the mathematical model, each training example is represented by an array or vector, sometimes called a feature vector, and the training data is represented by a matrix. Through iterative optimisation of an objective function, supervised learning algorithms learn a function that can be used to predict the output associated with new inputs.[52] An optimal function allows the algorithm to correctly determine the output for inputs that were not a part of the training data. An algorithm that improves the accuracy of its outputs or predictions over time is said to have learned to perform that task.[16]

    Types of supervised-learning algorithms include active learning, classification and regression.[53] Classification algorithms are used when the outputs are restricted to a limited set of values, while regression algorithms are used when the outputs can take any numerical value within a range. For example, in a classification algorithm that filters emails, the input is an incoming email, and the output is the folder in which to file the email. In contrast, regression is used for tasks such as predicting a person’s height based on factors like age and genetics or forecasting future temperatures based on historical data.[54]

    Similarity learning is an area of supervised machine learning closely related to regression and classification, but the goal is to learn from examples using a similarity function that measures how similar or related two objects are. It has applications in ranking, recommendation systems, visual identity tracking, face verification, and speaker verification.

    Unsupervised learning
    Main article: Unsupervised learning
    See also: Cluster analysis
    Unsupervised learning algorithms find structures in data that has not been labelled, classified or categorised. Instead of responding to feedback, unsupervised learning algorithms identify commonalities in the data and react based on the presence or absence of such commonalities in each new piece of data. Central applications of unsupervised machine learning include clustering, dimensionality reduction,[4] and density estimation.[55]

    Cluster analysis is the assignment of a set of observations into subsets (called clusters) so that observations within the same cluster are similar according to one or more predesignated criteria, while observations drawn from different clusters are dissimilar. Different clustering techniques make different assumptions on the structure of the data, often defined by some similarity metric and evaluated, for example, by internal compactness, or the similarity between members of the same cluster, and separation, the difference between clusters. Other methods are based on estimated density and graph connectivity.

    A special type of unsupervised learning called self-supervised learning involves training a model by generating the supervisory signal from the data itself.[56][57]

    Dimensionality reduction
    Dimensionality reduction is a process of reducing the number of random variables under consideration by obtaining a set of principal variables.[58] In other words, it is a process of reducing the dimension of the feature set, also called the “number of features”. Most of the dimensionality reduction techniques can be considered as either feature elimination or extraction. One of the popular methods of dimensionality reduction is principal component analysis (PCA). PCA involves changing higher-dimensional data (e.g., 3D) to a smaller space (e.g., 2D). The manifold hypothesis proposes that high-dimensional data sets lie along low-dimensional manifolds, and many dimensionality reduction techniques make this assumption, leading to the areas of manifold learning and manifold regularisation.

    Other dimensionality reduction techniques include t-distributed stochastic neighbor embedding (t-SNE) and the Self-organizing map (SOM) technique.

    Semi-supervised learning
    Main article: Semi-supervised learning
    Semi-supervised learning falls between unsupervised learning (without any labelled training data) and supervised learning (with completely labelled training data). Some of the training examples are missing training labels, yet many machine-learning researchers have found that unlabelled data, when used in conjunction with a small amount of labelled data, can produce a considerable improvement in learning accuracy.

    In weakly supervised learning, the training labels are noisy, limited, or imprecise; however, these labels are often cheaper to obtain, resulting in larger effective training sets.[59]

    Reinforcement learning
    Main article: Reinforcement learning

    In reinforcement learning, an agent takes actions in an environment: these produce a reward or a representation of the state, which is fed back to the agent.
    Reinforcement learning is an area of machine learning concerned with how software agents ought to take actions in an environment to maximise some notion of cumulative reward. Due to its generality, the field is studied in many other disciplines, such as game theory, control theory, operations research, information theory, simulation-based optimisation, multi-agent systems, swarm intelligence, statistics and genetic algorithms. In reinforcement learning, the environment is typically represented as a Markov decision process (MDP). Many reinforcement learning algorithms use dynamic programming techniques.[60] Reinforcement learning algorithms do not assume knowledge of an exact mathematical model of the MDP and are used when exact models are infeasible. Reinforcement learning algorithms are used in autonomous vehicles or in learning to play a game against a human opponent.

    Other types
    Other approaches have been developed which do not fit neatly into this three-fold categorisation, and sometimes more than one is used by the same machine learning system. For example, topic modelling, meta-learning.[61]

    Self-learning
    Self-learning, as a machine learning paradigm, was introduced in 1982 along with a neural network capable of self-learning, named crossbar adaptive array (CAA).[62][63] It gives a solution to the problem learning without any external reward, by introducing emotion as an internal reward. Emotion is used as a state evaluation of a self-learning agent. The CAA self-learning algorithm computes, in a crossbar fashion, both decisions about actions and emotions (feelings) about consequence situations. The system is driven by the interaction between cognition and emotion.[64] The self-learning algorithm updates a memory matrix W =||w(a,s)|| such that in each iteration executes the following machine learning routine:

    in situation s act a
    receive a consequence situation s’
    compute emotion of being in the consequence situation v(s’)
    update crossbar memory w'(a,s) = w(a,s) + v(s’)
    It is a system with only one input, situation, and only one output, action (or behaviour) a. There is neither a separate reinforcement input nor an advice input from the environment. The backpropagated value (secondary reinforcement) is the emotion toward the consequence situation. The CAA exists in two environments, one is the behavioural environment where it behaves, and the other is the genetic environment, wherefrom it initially and only once receives initial emotions about situations to be encountered in the behavioural environment. After receiving the genome (species) vector from the genetic environment, the CAA learns a goal-seeking behaviour in an environment that contains both desirable and undesirable situations.[65]

    Feature learning
    Main article: Feature learning
    Several learning algorithms aim at discovering better representations of the inputs provided during training.[66] Classic examples include principal component analysis and cluster analysis. Feature learning algorithms, also called representation learning algorithms, often attempt to preserve the information in their input but also transform it in a way that makes it useful, often as a pre-processing step before performing classification or predictions. This technique allows reconstruction of the inputs coming from the unknown data-generating distribution, while not being necessarily faithful to configurations that are implausible under that distribution. This replaces manual feature engineering, and allows a machine to both learn the features and use them to perform a specific task.

    Feature learning can be either supervised or unsupervised. In supervised feature learning, features are learned using labelled input data. Examples include artificial neural networks, multilayer perceptrons, and supervised dictionary learning. In unsupervised feature learning, features are learned with unlabelled input data. Examples include dictionary learning, independent component analysis, autoencoders, matrix factorisation[67] and various forms of clustering.[68][69][70]

    Manifold learning algorithms attempt to do so under the constraint that the learned representation is low-dimensional. Sparse coding algorithms attempt to do so under the constraint that the learned representation is sparse, meaning that the mathematical model has many zeros. Multilinear subspace learning algorithms aim to learn low-dimensional representations directly from tensor representations for multidimensional data, without reshaping them into higher-dimensional vectors.[71] Deep learning algorithms discover multiple levels of representation, or a hierarchy of features, with higher-level, more abstract features defined in terms of (or generating) lower-level features. It has been argued that an intelligent machine learns a representation that disentangles the underlying factors of variation that explain the observed data.[72]

    Feature learning is motivated by the fact that machine learning tasks such as classification often require input that is mathematically and computationally convenient to process. However, real-world data such as images, video, and sensory data have not yielded attempts to algorithmically define specific features. An alternative is to discover such features or representations through examination, without relying on explicit algorithms.

    Sparse dictionary learning
    Main article: Sparse dictionary learning
    Sparse dictionary learning is a feature learning method where a training example is represented as a linear combination of basis functions and assumed to be a sparse matrix. The method is strongly NP-hard and difficult to solve approximately.[73] A popular heuristic method for sparse dictionary learning is the k-SVD algorithm. Sparse dictionary learning has been applied in several contexts. In classification, the problem is to determine the class to which a previously unseen training example belongs. For a dictionary where each class has already been built, a new training example is associated with the class that is best sparsely represented by the corresponding dictionary. Sparse dictionary learning has also been applied in image denoising. The key idea is that a clean image patch can be sparsely represented by an image dictionary, but the noise cannot.[74]

    Anomaly detection
    Main article: Anomaly detection
    In data mining, anomaly detection, also known as outlier detection, is the identification of rare items, events or observations that raise suspicions by differing significantly from the majority of the data.[75] Typically, the anomalous items represent an issue such as bank fraud, a structural defect, medical problems or errors in a text. Anomalies are referred to as outliers, novelties, noise, deviations and exceptions.[76]

    In particular, in the context of abuse and network intrusion detection, the interesting objects are often not rare, but unexpected bursts of inactivity. This pattern does not adhere to the common statistical definition of an outlier as a rare object. Many outlier detection methods (in particular, unsupervised algorithms) will fail on such data unless aggregated appropriately. Instead, a cluster analysis algorithm may be able to detect the micro-clusters formed by these patterns.[77]

    Three broad categories of anomaly detection techniques exist.[78] Unsupervised anomaly detection techniques detect anomalies in an unlabelled test data set under the assumption that the majority of the instances in the data set are normal, by looking for instances that seem to fit the least to the remainder of the data set. Supervised anomaly detection techniques require a data set that has been labelled as “normal” and “abnormal” and involves training a classifier (the key difference from many other statistical classification problems is the inherently unbalanced nature of outlier detection). Semi-supervised anomaly detection techniques construct a model representing normal behaviour from a given normal training data set and then test the likelihood of a test instance being generated by the model.

    Robot learning
    Robot learning is inspired by a multitude of machine learning methods, starting from supervised learning, reinforcement learning,[79][80] and finally meta-learning (e.g. MAML).

    Association rules
    Main article: Association rule learning
    See also: Inductive logic programming
    Association rule learning is a rule-based machine learning method for discovering relationships between variables in large databases. It is intended to identify strong rules discovered in databases using some measure of “interestingness”.[81]

    Rule-based machine learning is a general term for any machine learning method that identifies, learns, or evolves “rules” to store, manipulate or apply knowledge. The defining characteristic of a rule-based machine learning algorithm is the identification and utilisation of a set of relational rules that collectively represent the knowledge captured by the system. This is in contrast to other machine learning algorithms that commonly identify a singular model that can be universally applied to any instance in order to make a prediction.[82] Rule-based machine learning approaches include learning classifier systems, association rule learning, and artificial immune systems.

    Based on the concept of strong rules, Rakesh Agrawal, Tomasz Imieliński and Arun Swami introduced association rules for discovering regularities between products in large-scale transaction data recorded by point-of-sale (POS) systems in supermarkets.[83] For example, the rule
    {
    o
    n
    i
    o
    n
    s
    ,
    p
    o
    t
    a
    t
    o
    e
    s
    }
    ⇒
    {
    b
    u
    r
    g
    e
    r
    }
    {\displaystyle {\mathrm {onions,potatoes} }\Rightarrow {\mathrm {burger} }} found in the sales data of a supermarket would indicate that if a customer buys onions and potatoes together, they are likely to also buy hamburger meat. Such information can be used as the basis for decisions about marketing activities such as promotional pricing or product placements. In addition to market basket analysis, association rules are employed today in application areas including Web usage mining, intrusion detection, continuous production, and bioinformatics. In contrast with sequence mining, association rule learning typically does not consider the order of items either within a transaction or across transactions.

    Learning classifier systems (LCS) are a family of rule-based machine learning algorithms that combine a discovery component, typically a genetic algorithm, with a learning component, performing either supervised learning, reinforcement learning, or unsupervised learning. They seek to identify a set of context-dependent rules that collectively store and apply knowledge in a piecewise manner to make predictions.[84]

    Inductive logic programming (ILP) is an approach to rule learning using logic programming as a uniform representation for input examples, background knowledge, and hypotheses. Given an encoding of the known background knowledge and a set of examples represented as a logical database of facts, an ILP system will derive a hypothesized logic program that entails all positive and no negative examples. Inductive programming is a related field that considers any kind of programming language for representing hypotheses (and not only logic programming), such as functional programs.

    Inductive logic programming is particularly useful in bioinformatics and natural language processing. Gordon Plotkin and Ehud Shapiro laid the initial theoretical foundation for inductive machine learning in a logical setting.[85][86][87] Shapiro built their first implementation (Model Inference System) in 1981: a Prolog program that inductively inferred logic programs from positive and negative examples.[88] The term inductive here refers to philosophical induction, suggesting a theory to explain observed facts, rather than mathematical induction, proving a property for all members of a well-ordered set.

    Models
    A machine learning model is a type of mathematical model that, once “trained” on a given dataset, can be used to make predictions or classifications on new data. During training, a learning algorithm iteratively adjusts the model’s internal parameters to minimise errors in its predictions.[89] By extension, the term “model” can refer to several levels of specificity, from a general class of models and their associated learning algorithms to a fully trained model with all its internal parameters tuned.[90]

    Various types of models have been used and researched for machine learning systems, picking the best model for a task is called model selection.

    Artificial neural networks
    Main article: Artificial neural network
    See also: Deep learning

    An artificial neural network is an interconnected group of nodes, akin to the vast network of neurons in a brain. Here, each circular node represents an artificial neuron and an arrow represents a connection from the output of one artificial neuron to the input of another.
    Artificial neural networks (ANNs), or connectionist systems, are computing systems vaguely inspired by the biological neural networks that constitute animal brains. Such systems “learn” to perform tasks by considering examples, generally without being programmed with any task-specific rules.

    An ANN is a model based on a collection of connected units or nodes called “artificial neurons”, which loosely model the neurons in a biological brain. Each connection, like the synapses in a biological brain, can transmit information, a “signal”, from one artificial neuron to another. An artificial neuron that receives a signal can process it and then signal additional artificial neurons connected to it. In common ANN implementations, the signal at a connection between artificial neurons is a real number, and the output of each artificial neuron is computed by some non-linear function of the sum of its inputs. The connections between artificial neurons are called “edges”. Artificial neurons and edges typically have a weight that adjusts as learning proceeds. The weight increases or decreases the strength of the signal at a connection. Artificial neurons may have a threshold such that the signal is only sent if the aggregate signal crosses that threshold. Typically, artificial neurons are aggregated into layers. Different layers may perform different kinds of transformations on their inputs. Signals travel from the first layer (the input layer) to the last layer (the output layer), possibly after traversing the layers multiple times.

    The original goal of the ANN approach was to solve problems in the same way that a human brain would. However, over time, ANNs have diverged from biology, as they have been optimized for engineering performance on specific tasks and implemented on digital hardware whose physical substrate and signaling mechanisms differ from those of biological neural tissue. These implementation-level differences may not preclude brain-inspired design at more abstract levels, including computational objectives, learning rules, network architectures[91][92], as well as broader principles, such as predictive or generative world models[93][94][95], biologically plausible credit assignment[96][97], and complementary fast- and slow-learning memory systems[98][99]. Artificial neural networks have been used on a variety of tasks, including computer vision, speech recognition, machine translation, social network filtering, playing board and video games and medical diagnosis.

    Deep learning consists of multiple hidden layers in an artificial neural network. This approach tries to model the way the human brain processes light and sound into vision and hearing. Some successful applications of deep learning are computer vision and speech recognition.[100]

    Decision trees
    Main article: Decision tree learning

    A decision tree showing survival probability of passengers on the Titanic
    Decision tree learning uses a decision tree as a predictive model to go from observations about an item (represented in the branches) to conclusions about the item’s target value (represented in the leaves). It is one of the predictive modelling approaches used in statistics, data mining, and machine learning. Tree models where the target variable can take a discrete set of values are called classification trees; in these tree structures, leaves represent class labels, and branches represent conjunctions of features that lead to those class labels. Decision trees where the target variable can take continuous values (typically real numbers) are called regression trees. In decision analysis, a decision tree can be used to visually and explicitly represent decisions and decision making. In data mining, a decision tree describes data, but the resulting classification tree can be an input for decision-making.

    Random forest regression
    Random forest regression (RFR) falls under the umbrella of decision tree-based models. RFR is an ensemble learning method that builds multiple decision trees and averages their predictions to improve accuracy and to avoid overfitting. To build decision trees, RFR uses bootstrapped sampling; for instance, each decision tree is trained on random data from the training set. This random selection of RFR for training enables the model to reduce biased predictions and achieve a higher degree of accuracy. RFR generates independent decision trees, and it can work on single-output data as well as multiple regressor tasks. This makes RFR compatible to be use in various applications.[101][102]

    Support-vector machines
    Main article: Support-vector machine
    Support-vector machines (SVMs), also known as support-vector networks, are a set of related supervised learning methods used for classification and regression. Given a set of training examples, each marked as belonging to one of two categories, an SVM training algorithm builds a model that predicts whether a new example falls into one category.[103] An SVM training algorithm is a non-probabilistic, binary, linear classifier, although methods such as Platt scaling exist to use SVM in a probabilistic classification setting. In addition to performing linear classification, SVMs can efficiently perform a non-linear classification using what is called the kernel trick, implicitly mapping their inputs into high-dimensional feature spaces.

    Regression analysis
    Main article: Regression analysis

    Illustration of linear regression on a data set
    Regression analysis encompasses a large variety of statistical methods to estimate the relationship between input variables and their associated features. Its most common form is linear regression, where a single line is drawn to best fit the given data according to a mathematical criterion such as ordinary least squares. The latter is often extended by regularisation methods to mitigate overfitting and bias, as in ridge regression. When dealing with non-linear problems, go-to models include polynomial regression (for example, used for trendline fitting in Microsoft Excel[104]), logistic regression (often used in statistical classification) or even kernel regression, which introduces non-linearity by taking advantage of the kernel trick to implicitly map input variables to higher-dimensional space.

    Multivariate linear regression extends the concept of linear regression to handle multiple dependent variables simultaneously. This approach estimates the relationships between a set of input variables and several output variables by fitting a multidimensional linear model. It is particularly useful in scenarios where outputs are interdependent or share underlying patterns, such as predicting multiple economic indicators or reconstructing images,[105] which are inherently multi-dimensional.

    Bayesian networks
    Main article: Bayesian network

    A simple Bayesian network. Rain influences whether the sprinkler is activated, and both rain and the sprinkler influence whether the grass is wet.
    A Bayesian network, belief network, or directed acyclic graphical model is a probabilistic graphical model that represents a set of random variables and their conditional independence with a directed acyclic graph (DAG). For example, a Bayesian network could represent the probabilistic relationships between diseases and symptoms. Given symptoms, the network can be used to compute the probabilities of the presence of various diseases. Efficient algorithms exist that perform inference and learning. Bayesian networks that model sequences of variables, like speech signals or protein sequences, are called dynamic Bayesian networks. Generalisations of Bayesian networks that can represent and solve decision problems under uncertainty are called influence diagrams.

    Gaussian processes
    Main article: Gaussian processes

    An example of Gaussian Process Regression (prediction) compared with other regression models[106]
    A Gaussian process is a stochastic process in which every finite collection of the random variables in the process has a multivariate normal distribution, and it relies on a pre-defined covariance function, or kernel, that models how pairs of points relate to each other depending on their locations.

    Given a set of observed points, or input–output examples, the distribution of the (unobserved) output of a new point as a function of its input data can be directly computed by looking at the observed points and the covariances between those points and the new, unobserved point.

    Gaussian processes are popular surrogate models in Bayesian optimisation used to do hyperparameter optimisation.

    Genetic algorithms
    Main article: Genetic algorithm
    A genetic algorithm (GA) is a search algorithm and heuristic technique that mimics the process of natural selection, using methods such as mutation and crossover to generate new genotypes in the hope of finding good solutions to a given problem. In machine learning, genetic algorithms were used in the 1980s and 1990s.[107][108] Conversely, machine learning techniques have been used to improve the performance of genetic and evolutionary algorithms.[109]

    Belief functions
    Main article: Dempster–Shafer theory
    The theory of belief functions, also referred to as evidence theory or Dempster–Shafer theory, is a general framework for reasoning with uncertainty, with understood connections to other frameworks such as probability, possibility and imprecise probability theories. These theoretical frameworks can be thought of as a kind of learner and have some analogous properties of how evidence is combined (e.g., Dempster’s rule of combination), just like how in a pmf-based Bayesian approach would combine probabilities.[110] However, there are many caveats to these beliefs functions when compared to Bayesian approaches to incorporate ignorance and uncertainty quantification. These belief function approaches that are implemented within the machine learning domain typically leverage a fusion approach of various ensemble methods to better handle the learner’s decision boundary, low samples, and ambiguous class issues that standard machine learning approach tend to have difficulty resolving.[111][7] However, the computational complexity of these algorithms is dependent on the number of propositions (classes), and can lead to a much higher computation time when compared to other machine learning approaches.

    Rule-based models
    Main article: Rule-based machine learning
    Rule-based machine learning (RBML) is a branch of machine learning that automatically discovers and learns ‘rules’ from data. It provides interpretable models, making it useful for decision-making in fields like healthcare, fraud detection, and cybersecurity. Key RBML techniques includes learning classifier systems,[112] association rule learning,[113] artificial immune systems,[114] and other similar models. These methods extract patterns from data and evolve rules over time.

    Training models
    Typically, machine learning models require a high quantity of reliable data to perform accurate predictions. When training a machine learning model, machine learning engineers need to target and collect a large and representative sample of data. Data from the training set can be as varied as a corpus of text, a collection of images, sensor data, and data collected from individual users of a service. Overfitting is something to watch out for when training a machine learning model. Trained models derived from biased or non-evaluated data can result in skewed or undesired predictions. Biased models may result in detrimental outcomes, thereby furthering the negative impacts on society or objectives. Algorithmic bias is a potential result of data not being fully prepared for training. Machine learning ethics is becoming a field of study and, notably, becoming integrated within machine learning engineering teams.

    Federated learning
    Main article: Federated learning
    Federated learning is an adapted form of distributed artificial intelligence to train machine learning models that decentralises the training process, allowing for users’ privacy to be maintained by not needing to send their data to a centralised server. This also increases efficiency by decentralising the training process to many devices. For example, Gboard uses federated machine learning to train search query prediction models on users’ mobile phones without having to send individual searches back to Google.[115]

    Applications
    There are many applications for machine learning, including:

    Agriculture
    Anatomy
    Adaptive website
    Affective computing
    Astronomy
    Automated decision-making
    Banking
    Behaviorism
    Bioinformatics
    Brain–machine interfaces
    Cheminformatics
    Citizen Science
    Climate Science
    Computer networks
    Computer vision
    Credit-card fraud detection
    Data quality
    DNA sequence classification
    Economics
    Financial data analysis[116]
    General game playing
    Handwriting recognition
    Healthcare
    Information retrieval
    Insurance
    Internet fraud detection
    Investment management[117]
    Knowledge graph embedding
    Linguistics
    Machine learning control
    Machine perception
    Machine translation
    Material Engineering
    Mathematics
    Marketing
    Medical diagnosis
    Natural language processing
    Natural language understanding
    Online advertising
    Optimisation
    Recommender systems
    Robot locomotion
    Search engines
    Sentiment analysis
    Sequence mining
    Software engineering
    Speech recognition
    Structural health monitoring
    Syntactic pattern recognition
    Telecommunications
    Theorem proving
    Time-series forecasting
    Tomographic reconstruction[118]
    User behaviour analytics
    In 2006, the media-services provider Netflix held the first “Netflix Prize” competition to find a program to better predict user preferences and improve the accuracy of its existing Cinematch movie recommendation algorithm by at least 10%. A joint team made up of researchers from AT&T Labs-Research in collaboration with the teams Big Chaos and Pragmatic Theory built an ensemble model to win the Grand Prize in 2009 for $1 million.[119] Shortly after the prize was awarded, Netflix realised that viewers’ ratings were not the best indicators of their viewing patterns (“everything is a recommendation”) and they changed their recommendation engine accordingly.[120] In 2010, an article in The Wall Street Journal noted the use of machine learning by Rebellion Research to predict the 2008 financial crisis.[121] In 2012, co-founder of Sun Microsystems, Vinod Khosla, predicted that 80% of medical doctors jobs would be lost in the next two decades to automated machine learning medical diagnostic software.[122] In 2014, it was reported that a machine learning algorithm had been applied in the field of art history to study fine art paintings and that it may have revealed previously unrecognised influences among artists.[123] In 2019 Springer Nature published the first research book created using machine learning.[124] In 2020, machine learning technology was used to help make diagnoses and aid researchers in developing a cure for COVID-19.[125] Machine learning was recently applied to predict the pro-environmental behaviour of travellers.[126] Recently, machine learning technology was also applied to optimise smartphone’s performance and thermal behaviour based on the user’s interaction with the phone.[127][128][129] When applied correctly, machine learning algorithms (MLAs) can utilise a wide range of company characteristics to predict stock returns without overfitting. By employing effective feature engineering and combining forecasts, MLAs can generate results that far surpass those obtained from basic linear techniques like OLS.[130]

    Recent advancements in machine learning have extended into the field of quantum chemistry, where novel algorithms now enable the prediction of solvent effects on chemical reactions, thereby offering new tools for chemists to tailor experimental conditions for optimal outcomes.[131]

    Machine Learning is becoming a useful tool to investigate and predict evacuation decision-making in large-scale and small-scale disasters. Different solutions have been tested to predict if and when householders decide to evacuate during wildfires and hurricanes.[132][133][134] Other applications have been focusing on pre evacuation decisions in building fires.[135][136]

    Limitations
    Although machine learning has been transformative in some fields, machine-learning programs often fail to deliver expected results.[137][138][139] Reasons for this are numerous: lack of (suitable) data, lack of access to the data, data bias, privacy problems, badly chosen tasks and algorithms, wrong tools and people, lack of resources, and evaluation problems.[140]

    The “black box theory” poses another yet significant challenge. Black box refers to a situation where the algorithm producing an output is entirely opaque, meaning that even the designers of an application cannot audit the pattern that the machine extracted from the data.[141] The House of Lords Select Committee, which claimed that such an “intelligence system” that could have a “substantial impact on an individual’s life” would not be considered acceptable unless it provided “a full and satisfactory explanation for the decisions” it makes.[141]

    In 2018, a self-driving car from Uber failed to detect a pedestrian, who was killed after a collision.[142] Attempts to use machine learning in healthcare with the IBM Watson system failed to deliver even after years of time and billions of dollars invested.[143][144] Microsoft’s Bing Chat chatbot has been reported to produce hostile and offensive response against its users.[145]

    Machine learning has been used as a strategy to update the evidence related to a systematic review and increased reviewer burden related to the growth of biomedical literature. While it has improved with training sets, it has not yet developed sufficiently to reduce the workload burden without limiting the necessary sensitivity for the findings research itself.[146]

    Explainability
    Main article: Explainable artificial intelligence
    Explainable AI (XAI), or Interpretable AI, or Explainable Machine Learning (XML), is artificial intelligence (AI) in which humans can understand the decisions or predictions made by the AI.[147] It contrasts with the “black box” concept in machine learning where even its designers cannot explain why an AI arrived at a specific decision.[148] By refining the mental models of users of AI-powered systems and dismantling their misconceptions, XAI promises to help users perform more effectively. XAI may be an implementation of the social right to explanation.

    Overfitting
    Main article: Overfitting

    The blue line could be an example of overfitting a linear function due to random noise.
    Settling on a bad, overly complex theory gerrymandered to fit all the past training data is known as overfitting. Many systems attempt to reduce overfitting by rewarding a theory in accordance with how well it fits the data but penalising the theory in accordance with how complex the theory is.[149]

    Model collapse
    This section is an excerpt from Model collapse.edit
    In artificial intelligence, model collapse, also known as “AI inbreeding”,[150][151] “AI cannibalism”,[152][153] “Habsburg AI”,[154] and “model autophagy disorder” or “MAD”,[155][156][157] is the degradation of machine learning models from uncurated synthetic data, or from training on the outputs of another model, such as a prior versions of itself. It has colloquially been referred to as the AI version of mad cow disease,[157][158] fundamentally caused by feeding something to itself. It is unclear to what extent the model collapse threatens the long-term development of AI models, and techniques have been proposed to mitigate the effect.

    Hallucinations
    This section is an excerpt from Hallucination (artificial intelligence).edit
    Duration: 9 seconds.0:09
    A video of the Glenfinnan Viaduct in Scotland generated by Sora, incorrectly showing: a second track, trains running on the right instead of the left, a second chimney on its interpretation of the train The Jacobite, inconsistent carriage lengths, unnatural amounts of visual noise, and a carriage bending along its length as it rounds the turn.

    The real Glenfinnan Viaduct with the real The Jacobite on it
    In the field of artificial intelligence (AI), a hallucination or artificial hallucination (also called bullshitting,[159][160] confabulation,[161] delusion,[162] or mirage[163]) is a response generated by AI that contains false or misleading information presented as fact.[164] These terms draw a loose analogy with human psychology, where a hallucination typically involves false percepts.

    Chatbots powered by large language models (LLMs), like ChatGPT, may embed plausible-sounding random falsehoods within their generated content, such as fabricated citations. Detecting and mitigating errors and hallucinations pose significant challenges for practical deployment and reliability of LLMs in high-stakes scenarios, such as chip design, supply chain logistics, and medical diagnostics.[165][166][167] Some software engineers and statisticians have criticized the specific term “AI hallucination” for unreasonably anthropomorphizing computers.[168][169] Symbolic artificial intelligence models generally do not produce hallucinations, unlike large language models.[170]

    Hallucinations are different from knowingly deceiving: a model may generate an incorrect statement without having an intention, belief, or awareness of whether it is true or false. These can happen when a generative model produces a statistically plausible continuation that isn’t supported by its input, training data, or outside information. Retrieval augmented generation, citation checking, constrained output formats, and human reviews are the best ways used to lower risk of hallucinated output, although none guarantee factual accuracy in every case.[171]

    Other limitations and vulnerabilities
    Learners can also be disappointed by “learning the wrong lesson”. A toy example is that an image classifier trained only on pictures of brown horses and black cats might conclude that all brown patches are likely to be horses.[172] A real-world example is that, unlike humans, current image classifiers often do not primarily make judgments from the spatial relationship between components of the picture, and they learn relationships between pixels that humans are oblivious to, but that still correlate with images of certain types of real objects. Modifying these patterns on a legitimate image can result in “adversarial” images that the system misclassifies.[173][174]

    Adversarial vulnerabilities can also result in nonlinear systems or from non-pattern perturbations. For some systems, it is possible to change the output by only changing a single adversarially chosen pixel.[175] Machine learning models are often vulnerable to manipulation or evasion via adversarial machine learning.[176]

    Researchers have demonstrated how backdoors can be placed undetectably into classifying (e.g., for categories “spam” and “not spam” of posts) machine learning models that are often developed or trained by third parties. Parties can change the classification of any input, including in cases for which a type of data/software transparency is provided, possibly including white-box access.[177][178][179]

    Model assessments
    Classification of machine learning models can be validated by accuracy estimation techniques like the holdout method, which splits the data into a training and test set (conventionally 2/3 training set and 1/3 test set designation) and evaluates the performance of the training model on the test set. In comparison, the K-fold-cross-validation method randomly partitions the data into K subsets and then K experiments are performed each considering 1 subset for evaluation and the remaining K-1 subsets for training the model. In addition to the holdout and cross-validation methods, bootstrap, which samples n instances with replacement from the dataset, can be used to assess model accuracy.[180]

    In addition to overall accuracy, investigators frequently report sensitivity and specificity, meaning true positive rate (TPR) and true negative rate (TNR), respectively. Similarly, investigators sometimes report the false positive rate (FPR) as well as the false negative rate (FNR). However, these rates are ratios that fail to reveal their numerators and denominators. Receiver operating characteristic (ROC), along with the accompanying Area Under the ROC Curve (AUC), offer additional tools for classification model assessment. Higher AUC is associated with a better performing model.[181]

    Ethics
    This section is an excerpt from Ethics of artificial intelligence.edit
    The ethics of artificial intelligence covers a broad range of topics within AI that are considered to have particular ethical stakes.[182] These include algorithmic biases, fairness, accountability, transparency, privacy, and regulation, particularly where AI systems influence or automate decisions affecting individuals or society.

    The field also addresses broader societal and technological challenges, including machine ethics (how to make machines that behave ethically), lethal autonomous weapon systems, arms race dynamics, AI safety and alignment, technological unemployment, AI-enabled misinformation,[183] how to treat certain AI systems if they have a moral status (AI welfare and rights), artificial superintelligence and existential risks.[182]

    Some application areas may also have particularly important ethical implications, like healthcare, education, criminal justice, or the military.

    Bias
    Main article: Algorithmic bias
    Different machine learning approaches can suffer from different data biases. A machine learning system trained specifically on current customers may not be able to predict the needs of new customer groups that are not represented in the training data. When trained on human-made data, machine learning is likely to pick up the constitutional and unconscious biases already present in society.[184]

    Systems that are trained on datasets collected with biases may exhibit these biases upon use (algorithmic bias), thus digitising cultural prejudices.[185] For example, in 1988, the UK’s Commission for Racial Equality found that St. George’s Medical School had been using a computer program trained from data of previous admissions staff and this program had denied nearly 60 candidates who were found to either be women or have non-European-sounding names.[184] Using job hiring data from a firm with racist hiring policies may lead to a machine learning system duplicating the bias by scoring job applicants by similarity to previous successful applicants.[186][187] Another example includes predictive policing company Geolitica’s predictive algorithm that resulted in “disproportionately high levels of over-policing in low-income and minority communities” after being trained with historical crime data.[188]

    While responsible collection of data and documentation of algorithmic rules used by a system is considered a critical part of machine learning, some researchers blame the lack of participation and representation of minority populations in the field of AI for machine learning’s vulnerability to biases.[189] In fact, according to research carried out by the Computing Research Association in 2021, “female faculty make up just 16.1%” of all faculty members who focus on AI among several universities around the world.[190] Furthermore, among the group of “new U.S. resident AI PhD graduates,” 45% identified as white, 22.4% as Asian, 3.2% as Hispanic, and 2.4% as African American, which further demonstrates a lack of diversity in the field of AI.[190]

    Language models learned from data have been shown to contain human-like biases.[191][192] Because human languages contain biases, machines trained on language corpora will necessarily also learn these biases.[193][194] In 2016, Microsoft tested Tay, a chatbot that learned from Twitter, and it quickly picked up racist and sexist language.[195]

    In an experiment carried out by ProPublica, a machine learning algorithm’s insight into the recidivism rates among prisoners falsely flagged “black defendants high risk twice as often as white defendants”.[188] In 2015, Google Photos once tagged a couple of black people as gorillas, which caused controversy. The gorilla label was subsequently removed, and in 2023, it still cannot recognise gorillas.[196] Similar issues with recognising non-white people have been found in many other systems.[197]

    Financial incentives
    There are concerns among health care professionals that these systems might not be designed in the public’s interest but as income-generating machines. This is especially true in the United States, where there is a long-standing ethical dilemma of improving health care, but also increasing profits. For example, the algorithms could be designed to provide patients with unnecessary tests or medication in which the algorithm’s proprietary owners hold stakes. There is potential for machine learning in health care to provide professionals with an additional tool to diagnose, medicate, and plan recovery paths for patients, but this requires these biases to be mitigated.[198]

    Hardware
    Since the 2010s, advances in both machine learning algorithms and computer hardware have led to more efficient methods for training deep neural networks (a particular narrow subdomain of machine learning) that contain many layers of nonlinear hidden units.[199] By 2019, graphics processing units (GPUs), often with AI-specific enhancements, had displaced CPUs as the dominant method of training large-scale commercial cloud AI.[200] OpenAI estimated the hardware compute used in the largest deep learning projects from AlexNet (2012) to AlphaZero (2017), and found a 300,000-fold increase in the amount of compute required, with a doubling-time trendline of 3.4 months.[201][202]

    Tensor Processing Units (TPUs)
    Tensor Processing Units (TPUs) are specialised hardware accelerators developed by Google specifically for machine learning workloads. Unlike general-purpose GPUs and FPGAs, TPUs are optimised for tensor computations, making them particularly efficient for deep learning tasks such as training and inference. They are widely used in Google Cloud AI services and large-scale machine learning models like Google’s DeepMind AlphaFold and large language models. TPUs leverage matrix multiplication units and high-bandwidth memory to accelerate computations while maintaining energy efficiency.[203] Since their introduction in 2016, TPUs have become a key component of AI infrastructure, especially in cloud-based environments.

    Neuromorphic computing
    Neuromorphic computing refers to a class of computing systems designed to emulate the structure and functionality of biological neural networks. These systems may be implemented through software-based simulations on conventional hardware or through specialised hardware architectures.[204]

    Physical neural networks
    A physical neural network is a specific type of neuromorphic hardware that relies on electrically adjustable materials, such as memristors, to emulate the function of neural synapses. The term “physical neural network” highlights the use of physical hardware for computation, as opposed to software-based implementations. It broadly refers to artificial neural networks that use materials with adjustable resistance to replicate neural synapses.[205][206]

    Embedded machine learning
    Embedded machine learning is a sub-field of machine learning where models are deployed on embedded systems with limited computing resources, such as wearable computers, edge devices and microcontrollers.[207][208][209][210] Running models directly on these devices eliminates the need to transfer and store data on cloud servers for further processing, thereby reducing the risk of data breaches, privacy leaks and theft of intellectual property, personal data and business secrets. Embedded machine learning can be achieved through various techniques, such as hardware acceleration,[211][212] approximate computing,[213] and model optimization.[214][215] Common optimization techniques include pruning, quantisation, knowledge distillation, low-rank factorisation, network architecture search, and parameter sharing

  • Deep learning

    For the TV series episode, see Deep Learning (South Park).
    Representing images on multiple layers of abstraction in deep learning
    Representing images on multiple layers of abstraction in deep learning[1]
    Part of a series on
    Artificial intelligence (AI)
    Major goals
    Approaches
    Machine learningData miningSymbolicDeep learningShallow learningBayesian networksEvolutionary algorithmsNeuro-symbolic AISystems integrationOpen-sourceOpen weightsAI data centers
    Applications
    Philosophy
    History
    Controversies
    Security Incidents
    Glossary
    vte
    In machine learning, deep learning (DL) focuses on utilizing multilayered neural networks to perform tasks such as classification, regression, and representation learning. The field takes inspiration from biological neuroscience and revolves around stacking artificial neurons into layers and “training” them to process data. The adjective “deep” refers to the use of multiple layers (ranging from three to several hundred or thousands) in the network. Methods used can be supervised, semi-supervised or unsupervised.[2]

    Some common deep learning network architectures include fully connected networks, deep belief networks, recurrent neural networks, recurrent Random neural networks, convolutional neural networks, generative adversarial networks, transformers, and neural radiance fields. These architectures have been applied to fields including computer vision, speech recognition, natural language processing, machine translation, bioinformatics, drug design, medical image analysis, climate science, material inspection and board game programs, where they have produced results comparable to and in some cases surpassing human expert performance.[3][4][5]

    Early forms of neural networks were inspired by information processing and distributed communication nodes in biological systems, particularly the human brain. However, current neural networks do not intend to model the brain function of organisms, and are generally seen as low-quality models for that purpose.[6]

    Overview
    Most modern deep learning models are based on multi-layered neural networks such as convolutional neural networks and transformers, although they can also include propositional formulas or latent variables organized layer-wise in deep generative models such as the nodes in deep belief networks and deep Boltzmann machines.[7]

    Fundamentally, deep learning refers to a class of machine learning algorithms in which a hierarchy of layers is used to transform input data into a progressively more abstract and composite representation. For example, in an image recognition model, the raw input may be an image (represented as a tensor of pixels). The first representational layer may attempt to identify basic shapes such as lines and circles, the second layer may compose and encode arrangements of edges, the third layer may encode a nose and eyes, and the fourth layer may recognize that the image contains a face.

    Importantly, a deep learning process can learn which features to optimally place at which level on its own. Prior to deep learning, machine learning techniques often involved hand-crafted feature engineering to transform the data into a more suitable representation for a classification algorithm to operate on. In the deep learning approach, features are not hand-crafted and the model discovers useful feature representations from the data automatically. This does not eliminate the need for hand-tuning; for example, varying numbers of layers and layer sizes can provide different degrees of abstraction.[8][2]

    The word “deep” in “deep learning” refers to the number of layers through which the data is transformed. More precisely, deep learning systems have a substantial credit assignment path (CAP) depth. The CAP is the chain of transformations from input to output. CAPs describe potentially causal connections between input and output. For a feedforward neural network, the depth of the CAPs is that of the network and is the number of hidden layers plus one (as the output layer is also parameterized). For recurrent neural networks, in which a signal may propagate through a layer more than once, the CAP depth is potentially unlimited.[9] No universally agreed-upon threshold of depth divides shallow learning from deep learning, but most researchers agree that deep learning involves CAP depth higher than two. CAP of depth two has been shown to be a universal approximator in the sense that it can emulate any function.[10][citation needed] Beyond that, more layers do not add to the function approximator ability of the network. Deep models (CAP > two) are able to extract better features than shallow models and hence, extra layers help in learning the features effectively.

    Deep learning architectures can be constructed with a greedy layer-by-layer method.[11] Deep learning helps to disentangle these abstractions and pick out which features improve performance.[8]

    Deep learning algorithms can be applied to unsupervised learning tasks. This is an important benefit because unlabeled data is more abundant than labeled data. Examples of deep structures that can be trained in an unsupervised manner are deep belief networks.[8][12]

    The term deep learning was introduced to the machine learning community by Rina Dechter in 1986,[13] and to artificial neural networks by Igor Aizenberg and colleagues in 2000, in the context of Boolean threshold neurons.[14][15] The etymology of the term is more complicated.[16]

    Interpretations
    Deep neural networks are generally interpreted in terms of the universal approximation theorem[17][18][19][20][21] or probabilistic inference.[22][23][8][9][24]

    The classic universal approximation theorem concerns the capacity of feedforward neural networks with a single hidden layer of finite size to approximate continuous functions.[17][18][19][20] In 1989, the first proof was published by George Cybenko for sigmoid activation functions[17] and was generalised to feed-forward multi-layer architectures in 1991 by Kurt Hornik.[18] Recent work also showed that universal approximation also holds for non-bounded activation functions such as Kunihiko Fukushima’s rectified linear unit.[25][26]. In [27] it was shown that the Random neural network is also a universal approximator for continuous and bounded functions.

    The universal approximation theorem for deep neural networks concerns the capacity of networks with bounded width but the depth is allowed to grow. Lu et al.[21] proved that if the width of a deep neural network with ReLU activation is strictly larger than the input dimension, then the network can approximate any Lebesgue integrable function; if the width is smaller or equal to the input dimension, then a deep neural network is not a universal approximator.

    The probabilistic interpretation[24] derives from the field of machine learning. It features inference,[23][7][8][9][12][24] as well as the optimization concepts of training and testing, related to fitting and generalization, respectively. More specifically, the probabilistic interpretation considers the activation nonlinearity as a cumulative distribution function.[24] The probabilistic interpretation led to the introduction of dropout as regularizer in neural networks. The probabilistic interpretation was introduced by researchers including Hopfield, Widrow and Narendra and popularized in surveys such as the one by Bishop.[28]

    History
    Before 1980
    There are two types of artificial neural network (ANN): feedforward neural network (FNN) or multilayer perceptron (MLP) and recurrent neural networks (RNN). RNNs have cycles in their connectivity structure, whereas FNNs do not. In the 1920s, Wilhelm Lenz and Ernst Ising created the Ising model[29][30] which is essentially a non-learning RNN architecture consisting of neuron-like threshold elements. In 1972, Shun’ichi Amari made this architecture adaptive.[31][32] His learning RNN was republished by John Hopfield in 1982.[33] Other early recurrent neural networks were published by Kaoru Nakano in 1971.[34][35] Already in 1948, Alan Turing produced work on “Intelligent Machinery” that was not published in his lifetime,[36] containing “ideas related to artificial evolution and learning RNNs”.[32]

    Frank Rosenblatt (1958)[37] proposed the perceptron, an MLP with 3 layers: an input layer, a hidden layer with randomized weights that did not learn, and an output layer. He later published a 1962 book that also introduced variants and computer experiments, including a version with four-layer perceptrons “with adaptive preterminal networks” where the last two layers have learned weights (here he credits H. D. Block and B. W. Knight).[38]: section 16  The book cites an earlier network by R. D. Joseph (1960)[39] “functionally equivalent to a variation of” this four-layer system (the book mentions Joseph over 30 times). Should Joseph therefore be considered the originator of proper adaptive multilayer perceptrons with learning hidden units? Unfortunately, the learning algorithm was not a functional one, and fell into oblivion.

    The first working deep learning algorithm was the Group method of data handling, a method to train arbitrarily deep neural networks, published by Alexey Ivakhnenko and Lapa in 1965. They regarded it as a form of polynomial regression,[40] or a generalization of Rosenblatt’s perceptron to handle more complex, nonlinear, and hierarchical relationships.[41] A 1971 paper described a deep network with eight layers trained by this method,[42] which is based on layer by layer training through regression analysis. Superfluous hidden units are pruned using a separate validation set. Since the activation functions of the nodes are Kolmogorov-Gabor polynomials, these were also the first deep networks with multiplicative units or “gates”.[32]

    The first deep learning multilayer perceptron trained by stochastic gradient descent[43] was published in 1967 by Shun’ichi Amari.[44] In computer experiments conducted by Amari’s student Saito, a five layer MLP with two modifiable layers learned internal representations to classify non-linearily separable pattern classes.[32] Subsequent developments in hardware and hyperparameter tunings have made end-to-end stochastic gradient descent the currently dominant training technique.

    In 1969, Kunihiko Fukushima introduced the ReLU (rectified linear unit) activation function.[25][32] The rectifier has become the most popular activation function for deep learning.[45]

    Deep learning architectures for convolutional neural networks (CNNs) with convolutional layers and downsampling layers began with the Neocognitron introduced by Kunihiko Fukushima in 1979, though not trained by backpropagation.[46][47]

    Backpropagation is an efficient application of the chain rule to networks of differentiable nodes. The terminology “back-propagating errors” was actually introduced in 1962 by Rosenblatt,[38] but he did not know how to implement this, although Henry J. Kelley had a continuous precursor of backpropagation in 1960 in the context of control theory.[48] The modern form of backpropagation was first published in Seppo Linnainmaa’s master thesis (1970).[49][50][32] G.M. Ostrovski et al. republished it in 1971.[51][52] Paul Werbos applied backpropagation to neural networks in 1982[53] (his 1974 PhD thesis, reprinted in a 1994 book,[54] did not yet describe the algorithm[52]). In 1986, David E. Rumelhart et al. popularised backpropagation but did not cite the original work.[55][56]

    1980s-2000s
    The time delay neural network (TDNN) was introduced in 1987 by Alex Waibel to apply CNN to phoneme recognition. It used convolutions, weight sharing, and backpropagation.[57][58] In 1988, Wei Zhang applied a backpropagation-trained CNN to alphabet recognition.[59] In 1989, Yann LeCun et al. created a CNN called LeNet for recognizing handwritten ZIP codes on mail. Training required 3 days.[60] In 1990, Wei Zhang implemented a CNN on optical computing hardware.[61] In 1991, a CNN was applied to medical image object segmentation[62] and breast cancer detection in mammograms.[63] LeNet-5 (1998), a 7-level CNN by Yann LeCun et al., that classifies digits, was applied by several banks to recognize hand-written numbers on checks digitized in 32×32 pixel images.[64]

    Recurrent neural networks (RNN)[29][31] were further developed in the 1980s. Recurrence is used for sequence processing, and when a recurrent network is unrolled, it mathematically resembles a deep feedforward layer. Consequently, they have similar properties and issues, and their developments had mutual influences. In RNN, two early influential works were the Jordan network (1986)[65], the Elman network (1990),[66] which applied RNN to study problems in cognitive psychology.

    The recurrent Random neural network introduced by Erol Gelenbe (1989) was inspired by the stochastic spiking behaviour of mammalian brain neurons, and [67] [68]. It was proved to have a convergent stable state that allowed the development of a simple deep learning gradient algorithm for this recurrent network [69]. It has had numerous applications, including for brain tumor detection from MR images using texture-based segmentation [70] and real-time video compression [71]. Gelenbe also developed its reinforcement learning algorithm, allowing the Random neural network to be used effectively for packet network routing that is interoperable and compatible with the existing Internet protocol [72] [73]. In recent years, new developments of the Random neural network that have large and dense clusters of recurrent neurons with Deep Learning, have resulted in highly accurate algorithms for cyberattack detection and mitigation [74] [75], as well as for the real-time control of smart homes.

    In the 1980s, backpropagation did not work well for deep learning with long credit assignment paths. To overcome this problem, in 1991, Jürgen Schmidhuber proposed a hierarchy of RNNs pre-trained one level at a time by self-supervised learning where each RNN tries to predict its own next input, which is the next unexpected input of the RNN below.[76][77] This “neural history compressor” uses predictive coding to learn internal representations at multiple self-organizing time scales. This can substantially facilitate downstream deep learning. The RNN hierarchy can be collapsed into a single RNN, by distilling a higher level chunker network into a lower level automatizer network.[76][77][32] In 1993, a neural history compressor solved a “Very Deep Learning” task that required more than 1000 subsequent layers in an RNN unfolded in time.[78] The “P” in ChatGPT refers to such pre-training.

    Sepp Hochreiter’s diploma thesis (1991)[79] implemented the neural history compressor,[76] and identified and analyzed the vanishing gradient problem.[79][80] Hochreiter proposed recurrent residual connections to solve the vanishing gradient problem. This led to the long short-term memory (LSTM), published in 1995.[81] LSTM can learn “very deep learning” tasks[9] with long credit assignment paths that require memories of events that happened thousands of discrete time steps before. That LSTM was not yet the modern architecture, which required a “forget gate”, introduced in 1999,[82] which became the standard RNN architecture.

    In 1991, Jürgen Schmidhuber also published adversarial neural networks that contest with each other in the form of a zero-sum game, where one network’s gain is the other network’s loss.[83][84] The first network is a generative model that models a probability distribution over output patterns. The second network learns by gradient descent to predict the reactions of the environment to these patterns. This was called “artificial curiosity”. In 2014, this principle was used in generative adversarial networks (GANs).[85]

    During 1985–1995, inspired by statistical mechanics, several architectures and methods were developed by Terry Sejnowski, Peter Dayan, Geoffrey Hinton, etc., including the Boltzmann machine,[86] restricted Boltzmann machine,[87] Helmholtz machine,[88] and the wake-sleep algorithm.[89] These were designed for unsupervised learning of deep generative models. However, those were more computationally expensive compared to backpropagation. Boltzmann machine learning algorithm, published in 1985, was briefly popular before being eclipsed by the backpropagation algorithm in 1986. (p. 112 [90]). A 1988 network became state of the art in protein structure prediction, an early application of deep learning to bioinformatics.[91]

    Both shallow and deep learning (e.g., recurrent nets) of ANNs for speech recognition have been explored for many years.[92][93][94] These methods never outperformed non-uniform internal-handcrafting Gaussian mixture model/Hidden Markov model (GMM-HMM) technology based on generative models of speech trained discriminatively.[95] Key difficulties have been analyzed, including gradient diminishing[79] and weak temporal correlation structure in neural predictive models.[96][97] Additional difficulties were the lack of training data and limited computing power.

    Most speech recognition researchers moved away from neural nets to pursue generative modeling. An exception was at SRI International in the late 1990s. Funded by the US government’s NSA and DARPA, SRI researched in speech and speaker recognition. The speaker recognition team led by Larry Heck reported significant success with deep neural networks in speech processing in the 1998 NIST Speaker Recognition benchmark.[98][99] It was deployed in the Nuance Verifier, representing the first major industrial application of deep learning.[100]

    The principle of elevating “raw” features over hand-crafted optimization was first explored successfully in the architecture of deep autoencoder on the “raw” spectrogram or linear filter-bank features in the late 1990s,[99] showing its superiority over the Mel-Cepstral features that contain stages of fixed transformation from spectrograms. The raw features of speech, waveforms, later produced excellent larger-scale results.[101]

    2000s
    Neural networks entered a lull, and simpler models that use task-specific handcrafted features such as Gabor filters and support vector machines (SVMs) became the preferred choices in the 1990s and 2000s, because of artificial neural networks’ computational cost and a lack of understanding of how the brain wires its biological networks.[citation needed]

    In 2003, LSTM became competitive with traditional speech recognizers on certain tasks.[102] In 2006, Alex Graves, Santiago Fernández, Faustino Gomez, and Schmidhuber combined it with connectionist temporal classification (CTC)[103] in stacks of LSTMs.[104] In 2009, it became the first RNN to win a pattern recognition contest, in connected handwriting recognition.[105][9]

    In 2006, publications by Geoff Hinton, Ruslan Salakhutdinov, Osindero and Teh[106][107] deep belief networks were developed for generative modeling. They are trained by training one restricted Boltzmann machine, then freezing it and training another one on top of the first one, and so on, then optionally fine-tuned using supervised backpropagation.[106] They could model high-dimensional probability distributions, such as the distribution of MNIST images, but convergence was slow.[106][108][109]

    The impact of deep learning in industry began in the early 2000s, when CNNs already processed an estimated 10% to 20% of all the checks written in the US, according to Yann LeCun.[110] Industrial applications of deep learning to large-scale speech recognition started around 2010.

    The 2009 NIPS Workshop on Deep Learning for Speech Recognition was motivated by the limitations of deep generative models of speech, and the possibility that given more capable hardware and large-scale data sets that deep neural nets might become practical. It was believed that pre-training DNNs using generative models of deep belief nets (DBN) would overcome the main difficulties of neural nets. However, it was discovered that replacing pre-training with large amounts of training data for straightforward backpropagation when using DNNs with large, context-dependent output layers produced error rates dramatically lower than then-state-of-the-art Gaussian mixture model (GMM)/Hidden Markov Model (HMM) and also than more-advanced generative model-based systems.[111] The nature of the recognition errors produced by the two types of systems was characteristically different,[112] offering technical insights into how to integrate deep learning into the existing highly efficient, run-time speech decoding system deployed by all major speech recognition systems.[23][113][114] Analysis around 2009–2010, contrasting the GMM (and other generative speech models) vs. DNN models, stimulated early industrial investment in deep learning for speech recognition.[112] That analysis was done with comparable performance (less than 1.5% in error rate) between discriminative DNNs and generative models.[111][112][115] In 2010, researchers extended deep learning from TIMIT to large vocabulary speech recognition, by adopting large output layers of the DNN based on context-dependent HMM states constructed by decision trees.[116][117][118][113]

    Deep learning revolution

    How deep learning is a subset of machine learning and how machine learning is a subset of artificial intelligence (AI)
    The period of significant growth in both the adoption and performance of deep learning in the 2010s was driven by the emergence of efficient implementations of CNNs on GPU hardware.

    Although CNNs trained by backpropagation had been around for decades and GPU implementations of NNs for years,[119] including CNNs,[120] faster implementations of CNNs on GPUs were needed to progress on computer vision. Later, as deep learning becomes widespread, specialized hardware and algorithm optimizations were developed specifically for deep learning.[121]

    A key advance for the deep learning revolution was hardware advances, especially GPU. Some early work dated back to 2004.[119][120] In 2009, Raina, Madhavan, and Andrew Ng reported a 100M deep belief network trained on 30 Nvidia GeForce GTX 280 GPUs, an early demonstration of GPU-based deep learning. They reported up to 70 times faster training.[122]

    In 2011, a CNN named DanNet[123][124] by Dan Ciresan, Ueli Meier, Jonathan Masci, Luca Maria Gambardella, and Jürgen Schmidhuber achieved for the first time superhuman performance in a visual pattern recognition contest, outperforming traditional methods by a factor of 3.[9] It then won more contests.[125][126] They also showed how max-pooling CNNs on GPU improved performance significantly.[3]

    In 2012, Andrew Ng and Jeff Dean created an FNN that learned to recognize higher-level concepts, such as cats, only from watching unlabeled images taken from YouTube videos.[127]

    In October 2012, AlexNet by Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton[4] won the large-scale ImageNet competition by a significant margin over shallow machine learning methods. Further incremental improvements included the VGG-16 network by Karen Simonyan and Andrew Zisserman[128] and Google’s Inceptionv3.[129]

    In 2013, Tomáš Mikolov and colleagues developed word2vec, a method for efficiently learning word embeddings from large text corpora.[130] The resulting vector representations were shown to capture syntactic and semantic relationships between words. A subsequent paper introduced negative sampling as an efficient method for training such models.[131]

    The success in image classification was then extended to the more challenging task of generating descriptions (captions) for images, often as a combination of CNNs and LSTMs.[132][133][134]

    In 2014, the state of the art was training a “very deep neural network” with 20 to 30 layers.[135] Stacking too many layers led to a steep reduction in training accuracy,[136] known as the “degradation” problem.[137] In 2015, two techniques were developed to train very deep networks: the highway network was published in May 2015, and the residual neural network (ResNet)[137] in Dec 2015. ResNet behaves like an open-gated Highway Net.

    Around the same time, deep learning started impacting the field of art. Early examples included Google DeepDream (2015), and neural style transfer (2015),[138] both of which were based on pretrained image classification neural networks, such as VGG-19.

    Generative adversarial network (GAN) by (Ian Goodfellow et al., 2014)[139] (based on Jürgen Schmidhuber’s principle of artificial curiosity[83][85]) became state of the art in generative modeling during 2014-2018 period. Excellent image quality is achieved by Nvidia’s StyleGAN (2018)[140] based on the Progressive GAN by Tero Karras et al.[141] Here the GAN generator is grown from small to large scale in a pyramidal fashion. Image generation by GAN reached popular success, and provoked discussions concerning deepfakes.[142] Diffusion models (2015)[143] eclipsed GANs in generative modeling since then, with systems such as DALL·E 2 (2022) and Stable Diffusion (2022).

    In 2015, Google’s speech recognition improved by 49% by an LSTM-based model, which they made available through Google Voice Search on smartphone.[144][145]

    Deep learning is part of state-of-the-art systems in various disciplines, particularly computer vision and automatic speech recognition (ASR). Results on commonly used evaluation sets such as TIMIT (ASR) and MNIST (image classification), as well as a range of large-vocabulary speech recognition tasks have steadily improved.[111][146] Convolutional neural networks were superseded for ASR by LSTM.[145][147][148][149] but are more successful in computer vision.

    Yoshua Bengio, Geoffrey Hinton and Yann LeCun were awarded the 2018 Turing Award for “conceptual and engineering breakthroughs that have made deep neural networks a critical component of computing”.[150]

    Neural networks
    Main article: Artificial neural network

    Simplified example of training a neural network in object detection: The network is trained by multiple images that are known to depict starfish and sea urchins, which are correlated with “nodes” that represent visual features. The starfish match with a ringed texture and a star outline, whereas most sea urchins match with a striped texture and oval shape. However, the instance of a ring textured sea urchin creates a weakly weighted association between them.

    Subsequent run of the network on an input image (left):[151] The network correctly detects the starfish. However, the weakly weighted association between ringed texture and sea urchin also confers a weak signal to the latter from one of two intermediate nodes. In addition, a shell that was not included in the training gives a weak signal for the oval shape, also resulting in a weak signal for the sea urchin output. These weak signals may result in a false positive result for sea urchin.
    In reality, textures and outlines would not be represented by single nodes, but rather by associated weight patterns of multiple nodes.
    Artificial neural networks (ANNs) or connectionist systems are computing systems inspired by the biological neural networks that constitute animal brains. Such systems learn (progressively improve their ability) to do tasks by considering examples, generally without task-specific programming. For example, in image recognition, they might learn to identify images that contain cats by analyzing example images that have been manually labeled as “cat” or “no cat” and using the analytic results to identify cats in other images. They have found most use in applications difficult to express with a traditional computer algorithm using rule-based programming.

    An ANN is based on a collection of connected units called artificial neurons, (analogous to biological neurons in a biological brain). Each connection (synapse) between neurons can transmit a signal to another neuron. The receiving (postsynaptic) neuron can process the signal(s) and then signal downstream neurons connected to it. Neurons may have state, generally represented by real numbers, typically between 0 and 1. Neurons and synapses may also have a weight that varies as learning proceeds, which can increase or decrease the strength of the signal that it sends downstream.

    Typically, neurons are organized in layers. Different layers may perform different kinds of transformations on their inputs. Signals travel from the first (input), to the last (output) layer, possibly after traversing the layers multiple times.

    The original goal of the neural network approach was to solve problems in the same way that a human brain would. Over time, attention focused on matching specific mental abilities, leading to deviations from biology such as backpropagation, or passing information in the reverse direction and adjusting the network to reflect that information.

    Neural networks have been used on a variety of tasks, including computer vision, speech recognition, machine translation, social network filtering, playing board and video games and medical diagnosis.

    As of 2017, neural networks typically have a few thousand to a few million units and millions of connections. Despite this number being several order of magnitude less than the number of neurons on a human brain, these networks can perform many tasks at a level beyond that of humans (e.g., recognizing faces, or playing “Go”[152]).

    Deep neural networks
    A deep neural network (DNN) is an artificial neural network with multiple layers between the input and output layers.[7][9] There are different types of neural networks but they always consist of the same components: neurons, synapses, weights, biases, and functions.[153] These components as a whole function in a way that mimics functions of the human brain, and can be trained like any other ML algorithm.[citation needed]

    For example, a DNN that is trained to recognize dog breeds will go over the given image and calculate the probability that the dog in the image is a certain breed. The user can review the results and select which probabilities the network should display (above a certain threshold, etc.) and return the proposed label. Each mathematical manipulation as such is considered a layer,[154] and complex DNN have many layers, hence the name “deep” networks.

    DNNs can model complex non-linear relationships. DNN architectures generate compositional models where the object is expressed as a layered composition of primitives.[155] The extra layers enable composition of features from lower layers, potentially modeling complex data with fewer units than a similarly performing shallow network.[7] For instance, it was proved that sparse multivariate polynomials are exponentially easier to approximate with DNNs than with shallow networks.[156]

    Deep architectures include many variants of a few basic approaches. Each architecture has found success in specific domains. It is not always possible to compare the performance of multiple architectures, unless they have been evaluated on the same data sets.[154]

    DNNs are typically feedforward networks in which data flows from the input layer to the output layer without looping back. At first, the DNN creates a map of virtual neurons and assigns random numerical values, or “weights”, to connections between them. The weights and inputs are multiplied and return an output between 0 and 1. If the network did not accurately recognize a particular pattern, an algorithm would adjust the weights.[157] That way the algorithm can make certain parameters more influential, until it determines the correct mathematical manipulation to fully process the data.

    Recurrent neural networks, in which data can flow in any direction, are used for applications such as language modeling.[158][159][160][161][162] Long short-term memory is particularly effective for this use.[163][164]

    Convolutional neural networks (CNNs) are used in computer vision.[165] CNNs also have been applied to acoustic modeling for automatic speech recognition (ASR).[166]

    Challenges
    As with ANNs, many issues can arise with naively trained DNNs. Two common issues are overfitting and computation time.

    DNNs are prone to overfitting because of the added layers of abstraction, which allow them to model rare dependencies in the training data. Regularization methods such as Ivakhnenko’s unit pruning[42] or weight decay (
    ℓ
    2
    {\displaystyle \ell _{2}}-regularization) or sparsity (
    ℓ
    1
    {\displaystyle \ell _{1}}-regularization) can be applied during training to combat overfitting.[167] Alternatively dropout regularization randomly omits units from the hidden layers during training. This helps to exclude rare dependencies.[168] Another interesting recent development is research into models of just enough complexity through an estimation of the intrinsic complexity of the task being modelled. This approach has been successfully applied for multivariate time series prediction tasks such as traffic prediction.[169] Finally, data can be augmented via methods such as cropping and rotating such that smaller training sets can be increased in size to reduce the chances of overfitting.[170]

    DNNs must consider many training parameters, such as the size (number of layers and number of units per layer), the learning rate, and initial weights. Sweeping through the parameter space for optimal parameters may not be feasible due to the cost in time and computational resources. Various tricks, such as batching (computing the gradient on several training examples at once rather than individual examples)[171] speed up computation. Large processing capabilities of many-core architectures (such as GPUs or the Intel Xeon Phi) have produced significant speedups in training, because of the suitability of such processing architectures for the matrix and vector computations.[172][173]

    Alternatively, engineers may look for other types of neural networks with more straightforward and convergent training algorithms. CMAC (cerebellar model articulation controller) is one such kind of neural network. It doesn’t require learning rates or randomized initial weights. The training process can be guaranteed to converge in one step with a new batch of data, and the computational complexity of the training algorithm is linear with respect to the number of neurons involved.[174][175]

    Hardware
    Since the 2010s, advances in both machine learning algorithms and computer hardware have led to more efficient methods for training deep neural networks that contain many layers of non-linear hidden units and a very large output layer.[176] By 2019, graphics processing units (GPUs), often with AI-specific enhancements, had displaced CPUs as the dominant method for training large-scale commercial cloud AI .[177] OpenAI estimated the hardware computation used in the largest deep learning projects from AlexNet (2012) to AlphaZero (2017) and found a 300,000-fold increase in the amount of computation required, with a doubling-time trendline of 3.4 months.[178][179]

    Special electronic circuits called deep learning processors were designed to speed up deep learning algorithms. Deep learning processors include neural processing units (NPUs) in Huawei cellphones[180] and cloud computing servers such as tensor processing units (TPU) in the Google Cloud Platform.[181] Cerebras Systems has also built a dedicated system to handle large deep learning models, the CS-2, based on the largest processor in the industry, the second-generation Wafer Scale Engine (WSE-2).[182][183]

    Atomically thin semiconductors are considered promising for energy-efficient deep learning hardware where the same basic device structure is used for both logic operations and data storage. In 2020, Marega et al. published experiments with a large-area active channel material for developing logic-in-memory devices and circuits based on floating-gate field-effect transistors (FGFETs).[184]

    In 2021, J. Feldmann et al. proposed an integrated photonic hardware accelerator for parallel convolutional processing.[185] The authors identify two key advantages of integrated photonics over its electronic counterparts: (1) massively parallel data transfer through wavelength division multiplexing in conjunction with frequency combs, and (2) extremely high data modulation speeds.[185] Their system can execute trillions of multiply-accumulate operations per second, indicating the potential of integrated photonics in data-heavy AI applications.[185]

    Applications
    Automatic speech recognition
    Main article: Speech recognition
    Large-scale automatic speech recognition is the first and most convincing successful case of deep learning. LSTM RNNs can learn “Very Deep Learning” tasks[9] that involve multi-second intervals containing speech events separated by thousands of discrete time steps, where one time step corresponds to about 10 ms. LSTM with forget gates[164] is competitive with traditional speech recognizers on certain tasks.[102]

    The initial success in speech recognition was based on small-scale recognition tasks based on TIMIT. The data set contains 630 speakers from eight major dialects of American English, where each speaker reads 10 sentences.[186] Its small size lets many configurations be tried. More importantly, the TIMIT task concerns phone-sequence recognition, which, unlike word-sequence recognition, allows weak phone bigram language models. This lets the strength of the acoustic modeling aspects of speech recognition be more easily analyzed. The error rates listed below, including these early results and measured as percent phone error rates (PER), have been summarized since 1991.

    Method Percent phone
    error rate (PER) (%)
    Randomly Initialized RNN[187] 26.1
    Bayesian Triphone GMM-HMM 25.6
    Hidden Trajectory (Generative) Model 24.8
    Monophone Randomly Initialized DNN 23.4
    Monophone DBN-DNN 22.4
    Triphone GMM-HMM with BMMI Training 21.7
    Monophone DBN-DNN on fbank 20.7
    Convolutional DNN[188] 20.0
    Convolutional DNN w. Heterogeneous Pooling 18.7
    Ensemble DNN/CNN/RNN[189] 18.3
    Bidirectional LSTM 17.8
    Hierarchical Convolutional Deep Maxout Network[190] 16.5
    The debut of DNNs for speaker recognition in the late 1990s and speech recognition around 2009-2011 and of LSTM around 2003–2007, accelerated progress in eight major areas:[23][115][113]

    Scale-up/out and accelerated DNN training and decoding
    Sequence discriminative training
    Feature processing by deep models with solid understanding of the underlying mechanisms
    Adaptation of DNNs and related deep models
    Multi-task and transfer learning by DNNs and related deep models
    CNNs and how to design them to best exploit domain knowledge of speech
    RNN and its rich LSTM variants
    Other types of deep models including tensor-based models and integrated deep generative/discriminative models.
    More recent speech recognition models use Transformers or Temporal Convolution Networks with significant success and widespread applications.[191][192][193] All major commercial speech recognition systems (e.g., Microsoft Cortana, Xbox, Skype Translator, Amazon Alexa, Google Now, Apple Siri, Baidu and iFlyTek voice search, and a range of Nuance speech products, etc.) are based on deep learning.[23][194][195]

    Image recognition
    Main article: Computer vision
    Richard Green explains how deep learning is used with a remotely operated vehicle in mussel aquaculture.
    A common evaluation set for image classification is the MNIST database data set. MNIST is composed of handwritten digits and includes 60,000 training examples and 10,000 test examples. As with TIMIT, its small size lets users test multiple configurations. A comprehensive list of results on this set is available.[196]

    Deep learning-based image recognition has become “superhuman”, producing more accurate results than human contestants. This first occurred in 2011 in recognition of traffic signs, and in 2014, with recognition of human faces.[197][198]

    Deep learning-trained vehicles now interpret 360° camera views.[199] Another example is Facial Dysmorphology Novel Analysis (FDNA) used to analyze cases of human malformation connected to a large database of genetic syndromes.

    Visual art processing

    Visual art processing of Jimmy Wales in France, with the style of Munch’s “The Scream” applied using neural style transfer
    Closely related to the progress that has been made in image recognition is the increasing application of deep learning techniques to various visual art tasks. DNNs have proven themselves capable, for example, of

    identifying the style period of a given painting[200][201]
    Neural Style Transfer – capturing the style of a given artwork and applying it in a visually pleasing manner to an arbitrary photograph or video[200][201]
    generating striking imagery based on random visual input fields.[200][201]
    Natural language processing
    Main article: Natural language processing
    Neural networks have been used for implementing language models since the early 2000s.[158] LSTM helped to improve machine translation and language modeling.[159][160][161]

    Other key techniques in this field are negative sampling[202] and word embedding. Word embedding, such as word2vec, can be thought of as a representational layer in a deep learning architecture that transforms an atomic word into a positional representation of the word relative to other words in the dataset; the position is represented as a point in a vector space. Using word embedding as an RNN input layer allows the network to parse sentences and phrases using an effective compositional vector grammar. A compositional vector grammar can be thought of as probabilistic context free grammar (PCFG) implemented by an RNN.[203] Recursive auto-encoders built atop word embeddings can assess sentence similarity and detect paraphrasing.[203] Deep neural architectures provide the best results for constituency parsing,[204] sentiment analysis,[205] information retrieval,[206][207] spoken language understanding,[208] machine translation,[159][209] contextual entity linking,[209] writing style recognition,[210] named-entity recognition (token classification),[211] text classification, and others.[212]

    Recent developments generalize word embedding to sentence embedding.

    Google Translate (GT) uses a large end-to-end long short-term memory (LSTM) network.[213][214][215][216] Google Neural Machine Translation (GNMT) uses an example-based machine translation method in which the system “learns from millions of examples”.[214] It translates “whole sentences at a time, rather than pieces”. Google Translate supports over one hundred languages.[214] The network encodes the “semantics of the sentence rather than simply memorizing phrase-to-phrase translations”.[214][217] GT uses English as an intermediate between most language pairs.[217]

    Drug discovery and toxicology
    For more information, see Drug discovery and Toxicology.
    A large percentage of candidate drugs fail to win regulatory approval. These failures are caused by insufficient efficacy (on-target effect), undesired interactions (off-target effects), or unanticipated toxic effects.[218][219] Research has explored use of deep learning to predict the biomolecular targets,[220][221] off-targets, and toxic effects of environmental chemicals in nutrients, household products and drugs.[222][223][224]

    AtomNet is a deep learning system for structure-based rational drug design.[225] AtomNet was used to predict novel candidate biomolecules for disease targets such as the Ebola virus[226] and multiple sclerosis.[227][226]

    In 2017 graph neural networks were used for the first time to predict various properties of molecules in a large toxicology data set.[228] In 2019, generative neural networks were used to produce molecules that were validated experimentally all the way into mice.[229][230]

    Recommendation systems
    Main article: Recommender system
    Recommendation systems have used deep learning to extract meaningful features for a latent factor model for content-based music and journal recommendations.[231][232] Multi-view deep learning has been applied for learning user preferences from multiple domains.[233] The model uses a hybrid collaborative and content-based approach and enhances recommendations in multiple tasks.

    Bioinformatics
    Main article: Bioinformatics
    An autoencoder ANN was used in bioinformatics, to predict gene ontology annotations and gene-function relationships.[234]

    In medical informatics, deep learning was used to predict sleep quality based on data from wearables[235] and predictions of health complications from electronic health record data.[236]

    Deep neural networks have shown unparalleled performance in predicting protein structure, according to the sequence of the amino acids that make it up. In 2020, AlphaFold, a deep-learning based system, achieved a level of accuracy significantly higher than all previous computational methods.[237][238]

    Deep Neural Network Estimations
    Deep neural networks can be used to estimate the entropy of a stochastic process through an arrangement called a Neural Joint Entropy Estimator (NJEE).[239] Such an estimation provides insights on the effects of input random variables on an independent random variable. Practically, the DNN is trained as a classifier that maps an input vector or matrix X to an output probability distribution over the possible classes of random variable Y, given input X. For example, in image classification tasks, the NJEE maps a vector of pixels’ color values to probabilities over possible image classes. In practice, the probability distribution of Y is obtained by a Softmax layer with number of nodes that is equal to the alphabet size of Y. NJEE uses continuously differentiable activation functions, such that the conditions for the universal approximation theorem holds. It is shown that this method provides a strongly consistent estimator and outperforms other methods in cases of large alphabet sizes.[239]

    Medical image analysis
    Deep learning has been shown to produce competitive results in medical applications such as cancer cell classification, lesion detection, organ segmentation and image enhancement.[240][241] Modern deep learning tools demonstrate the high accuracy of detecting various diseases and the helpfulness of their use by specialists to improve the diagnosis efficiency.[242][243]

    Mobile advertising
    Finding the appropriate mobile audience for mobile advertising is always challenging, since many data points must be considered and analyzed before a target segment can be created and used in ad serving by any ad server.[244] Deep learning has been used to interpret large, many-dimensioned advertising datasets. Many data points are collected during the request/serve/click internet advertising cycle. This information can form the basis of machine learning to improve ad selection.

    Image restoration
    Deep learning has been successfully applied to inverse problems such as denoising, super-resolution, inpainting, and film colorization.[245] These applications include learning methods such as “Shrinkage Fields for Effective Image Restoration”[246] which trains on an image dataset, and Deep Image Prior, which trains on the image that needs restoration.

    Financial fraud detection
    Deep learning is being applied to financial fraud detection, tax evasion detection,[247] and anti-money laundering.[248]

    Materials science
    In November 2023, researchers at Google DeepMind and Lawrence Berkeley National Laboratory announced that they had developed an AI system known as GNoME. This system has contributed to materials science by discovering over 2 million new materials within a relatively short timeframe. GNoME employs deep learning techniques to efficiently explore potential material structures, achieving a significant increase in the identification of stable inorganic crystal structures. The system’s predictions were validated through autonomous robotic experiments, demonstrating a noteworthy success rate of 71%. The data of newly discovered materials is publicly available through the Materials Project database, offering researchers the opportunity to identify materials with desired properties for various applications. This development has implications for the future of scientific discovery and the integration of AI in material science research, potentially expediting material innovation and reducing costs in product development. The use of AI and deep learning suggests the possibility of minimizing or eliminating manual lab experiments and allowing scientists to focus more on the design and analysis of unique compounds.[249][250][251]

    Military
    The United States Department of Defense applied deep learning to train robots in new tasks through observation.[252]

    Partial differential equations
    Physics-informed neural networks have been used to solve partial differential equations in both forward and inverse problems in a data driven manner.[253] One example is the reconstructing fluid flow governed by the Navier-Stokes equations. Using physics informed neural networks does not require the often expensive mesh generation that conventional CFD methods rely on.[254][255] It is evident that geometric and physical constraints have a synergistic effect on neural PDE surrogates, thereby enhancing their efficacy in predicting stable and super long rollouts.[256]

    Deep backward stochastic differential equation method
    Deep backward stochastic differential equation method is a numerical method that combines deep learning with Backward stochastic differential equation (BSDE). This method is particularly useful for solving high-dimensional problems in financial mathematics. By leveraging the powerful function approximation capabilities of deep neural networks, deep BSDE addresses the computational challenges faced by traditional numerical methods in high-dimensional settings. Specifically, traditional methods like finite difference methods or Monte Carlo simulations often struggle with the curse of dimensionality, where computational cost increases exponentially with the number of dimensions. Deep BSDE methods, however, employ deep neural networks to approximate solutions of high-dimensional partial differential equations (PDEs), effectively reducing the computational burden.[257]

    In addition, the integration of Physics-informed neural networks (PINNs) into the deep BSDE framework enhances its capability by embedding the underlying physical laws directly into the neural network architecture. This ensures that the solutions not only fit the data but also adhere to the governing stochastic differential equations. PINNs leverage the power of deep learning while respecting the constraints imposed by the physical models, resulting in more accurate and reliable solutions for financial mathematics problems.

    Image reconstruction
    Image reconstruction is the reconstruction of the underlying images from the image-related measurements. Several works showed the better and superior performance of the deep learning methods compared to analytical methods for various applications, e.g., spectral imaging [258] and ultrasound imaging.[259]

    Weather prediction
    Traditional weather prediction systems typically solve a complex system of partial differential equations. GraphCast is a deep learning-based model trained on a long history of weather data to predict how weather patterns change over time. It can predict weather conditions for up to 10 days globally, at a very detailed level, and in under a minute, with precision similar to state of the art systems.[260][261]

    Epigenetic clock
    Main article: Epigenetic clock
    An epigenetic clock is a biochemical test that can be used to measure age. Galkin et al. used deep neural networks to train an epigenetic aging clock of unprecedented accuracy using >6,000 blood samples.[262] The clock uses information from 1000 CpG sites and predicts people with certain conditions older than healthy controls: IBD, frontotemporal dementia, ovarian cancer, obesity. The aging clock was planned to be released for public use in 2021 by an Insilico Medicine spinoff company Deep Longevity.

    Relation to human cognitive and brain development
    Deep learning is closely related to a class of theories of brain development (specifically, neocortical development) proposed by cognitive neuroscientists in the early 1990s.[263][264][265][266] These developmental theories were instantiated in computational models, making them predecessors of deep learning systems. These developmental models share the property that various proposed learning dynamics in the brain (e.g., a wave of nerve growth factor) support the self-organization somewhat analogous to the neural networks utilized in deep learning models. Like the neocortex, neural networks employ a hierarchy of layered filters in which each layer considers information from a prior layer (or the operating environment), and then passes its output (and possibly the original input), to other layers. This process yields a self-organizing stack of transducers, well-tuned to their operating environment. A 1995 description stated, “…the infant’s brain seems to organize itself under the influence of waves of so-called trophic-factors … different regions of the brain become connected sequentially, with one layer of tissue maturing before another and so on until the whole brain is mature”.[267]

    A variety of approaches have been used to investigate the plausibility of deep learning models from a neurobiological perspective. On the one hand, several variants of the backpropagation algorithm have been proposed in order to increase its processing realism.[268][269] Other researchers have argued that unsupervised forms of deep learning, such as those based on hierarchical generative models and deep belief networks, may be closer to biological reality.[270][271] In this respect, generative neural network models have been related to neurobiological evidence about sampling-based processing in the cerebral cortex.[272]

    Although a systematic comparison between the human brain organization and the neuronal encoding in deep networks has not yet been established, several analogies have been reported. For example, the computations performed by deep learning units could be similar to those of actual neurons[273] and neural populations.[274] Similarly, the representations developed by deep learning models are similar to those measured in the primate visual system[275] both at the single-unit[276] and at the population[277] levels.

    Commercial activity
    Facebook’s AI lab performs tasks such as automatically tagging uploaded pictures with the names of the people in them.[278]

    Google’s DeepMind Technologies developed a system capable of learning how to play Atari video games using only pixels as data input. In 2015 they demonstrated their AlphaGo system, which learned the game of Go well enough to beat a professional Go player.[279][280][281] Google Translate uses a neural network to translate between more than 100 languages.

    In 2017, Covariant.ai was launched, which focuses on integrating deep learning into factories.[282]

    As of 2008,[283] researchers at The University of Texas at Austin (UT) developed a machine learning framework called Training an Agent Manually via Evaluative Reinforcement, or TAMER, which proposed new methods for robots or computer programs to learn how to perform tasks by interacting with a human instructor.[252] First developed as TAMER, a new algorithm called Deep TAMER was later introduced in 2018 during a collaboration between U.S. Army Research Laboratory (ARL) and UT researchers. Deep TAMER used deep learning to provide a robot with the ability to learn new tasks through observation.[252] Using Deep TAMER, a robot learned a task with a human trainer, watching video streams or observing a human perform a task in-person. The robot later practiced the task with the help of some coaching from the trainer, who provided feedback such as “good job” and “bad job”.[284]

    Criticism and comment
    Deep learning has attracted both criticism and comment, in some cases from outside the field of computer science.

    Theory
    See also: Explainable artificial intelligence
    A main criticism concerns the lack of theory surrounding some methods.[285] Learning in the most common deep architectures is implemented using well-understood gradient descent. However, the theory surrounding other algorithms, such as contrastive divergence is less clear.[citation needed] (e.g., Does it converge? If so, how fast? What is it approximating?) Deep learning methods are often looked at as a black box, with most confirmations done empirically, rather than theoretically.[286]

    In further reference to the idea that artistic sensitivity might be inherent in relatively low levels of the cognitive hierarchy, a published series of graphic representations of the internal states of deep (20-30 layers) neural networks attempting to discern within essentially random data the images on which they were trained[287] demonstrate a visual appeal: the original research notice received well over 1,000 comments, and was the subject of what was for a time the most frequently accessed article on The Guardian’s[288] website.

    With the support of Innovation Diffusion Theory (IDT), a study analyzed the diffusion of deep learning in BRICS and OECD countries using data from Google Trends.[289]

    Errors
    Some deep learning architectures display problematic behaviors,[290] such as confidently classifying unrecognizable images as belonging to a familiar category of ordinary images (2014)[291] and misclassifying minuscule perturbations of correctly classified images (2013).[292] Goertzel hypothesized that these behaviors are due to limitations in their internal representations and that these limitations would inhibit integration into heterogeneous multi-component artificial general intelligence (AGI) architectures.[290] These issues may possibly be addressed by deep learning architectures that internally form states homologous to image-grammar[293] decompositions of observed entities and events.[290] Learning a grammar (visual or linguistic) from training data would be equivalent to restricting the system to commonsense reasoning that operates on concepts in terms of grammatical production rules and is a basic goal of both human language acquisition[294] and artificial intelligence (AI).[295]

    Cyber threat
    As deep learning moves from the lab into the world, research and experience show that artificial neural networks are vulnerable to hacks and deception.[296] By identifying patterns that these systems use to function, attackers can modify inputs to ANNs in such a way that the ANN finds a match that human observers would not recognize. For example, an attacker can make subtle changes to an image such that the ANN finds a match even though the image looks to a human nothing like the search target. Such manipulation is termed an “adversarial attack”.[297]

    In 2016 researchers used one ANN to doctor images in trial and error fashion, identify another’s focal points, and thereby generate images that deceived it. The modified images looked no different to human eyes. Another group showed that printouts of doctored images then photographed successfully tricked an image classification system.[298] One defense is reverse image search, in which a possible fake image is submitted to a site such as TinEye that can then find other instances of it. A refinement is to search using only parts of the image, to identify images from which that piece may have been taken.[299]

    Another group showed that certain psychedelic spectacles could fool a facial recognition system into thinking ordinary people were celebrities, potentially allowing one person to impersonate another. In 2017 researchers added stickers to stop signs and caused an ANN to misclassify them.[298]

    ANNs can however be further trained to detect attempts at deception, potentially leading attackers and defenders into an arms race similar to the kind that already defines the malware defense industry. ANNs have been trained to defeat ANN-based anti-malware software by repeatedly attacking a defense with malware that was continually altered by a genetic algorithm until it tricked the anti-malware while retaining its ability to damage the target.[298]

    In 2016, another group demonstrated that certain sounds could make the Google Now voice command system open a particular web address, and hypothesized that this could “serve as a stepping stone for further attacks (e.g., opening a web page hosting drive-by malware)”.[298]

    In “data poisoning”, false data is continually smuggled into a machine learning system’s training set to prevent it from achieving mastery.[298]

  • Cloud computing

    “Cloud Computing” redirects here; not to be confused with Cloud Computing (horse).

    This article appears to incorporate text from a large language model, which is prohibited in Wikipedia articles. Any such material should be removed. The reason given is: various additions throughout 2025-26; note WP:AISIGNS – superficial analyses, vocab distro typical of LLM output, etc (April 2026) (Learn how and when to remove this message)

    Cloud computing metaphor: the group of networked elements providing services does not need to be addressed or managed individually by users; instead, the entire provider-managed suite of hardware and software can be thought of as an amorphous cloud.
    Cloud computing is defined by the International Organization for Standardization (ISO) as “a paradigm for enabling network access to a scalable and elastic pool of shareable physical or virtual resources with self-service provisioning and administration on demand”.[1] It is commonly referred to as “the cloud”.[2]

    Characteristics
    In 2011, the National Institute of Standards and Technology (NIST) identified five “essential characteristics” for cloud systems.[3] Below are the exact definitions according to NIST:[3]

    On-demand self-service: “A consumer can unilaterally provision computing capabilities, such as server time and network storage, as needed automatically without requiring human interaction with each service provider.”
    Broad network access: “Capabilities are available over the network and accessed through standard mechanisms that promote use by heterogeneous thin or thick client platforms (e.g., mobile phones, tablets, laptops, and workstations).”
    Resource pooling: ” The provider’s computing resources are pooled to serve multiple consumers using a multi-tenant model, with different physical and virtual resources dynamically assigned and reassigned according to consumer demand.”
    Rapid elasticity: “Capabilities can be elastically provisioned and released, in some cases automatically, to scale rapidly outward and inward commensurate with demand. To the consumer, the capabilities available for provisioning often appear unlimited and can be appropriated in any quantity at any time.”
    Measured service: “Cloud systems automatically control and optimize resource use by leveraging a metering capability at some level of abstraction appropriate to the type of service (e.g., storage, processing, bandwidth, and active user accounts). Resource usage can be monitored, controlled, and reported, providing transparency for both the provider and consumer of the utilized service.
    By 2023, the International Organization for Standardization (ISO) had expanded and refined the list.[4]

    History
    Main article: History of cloud computing
    The history of cloud computing extends to the 1960s, with the initial concepts of time-sharing becoming popularized via remote job entry (RJE). The “data center” model, where users submitted jobs to operators to run on mainframes, was predominantly used during this era. This period saw broad experimentation with making large-scale computing power more accessible through time-sharing, while optimizing infrastructure, platforms, and applications to improve efficiency for end users.[5]

    The “cloud” metaphor for virtualized services dates to 1994, when it was used by General Magic for the universe of “places” that mobile agents in the Telescript environment could “go”. The metaphor is credited to David Hoffman, a General Magic communications specialist, based on its long-standing use in networking and telecom.[6] The expression cloud computing became more widely known in 1996 when Compaq Computer Corporation drew up a business plan for future computing and the Internet. The company’s ambition was to supercharge sales with “cloud computing-enabled applications”. The business plan foresaw that online consumer file storage would likely be commercially successful. As a result, Compaq decided to sell server hardware to internet service providers.[7]

    In the 2000s, the application of cloud computing began to take shape with the establishment of Amazon Web Services (AWS) in 2002, which allowed developers to build applications independently. In 2006 Amazon Simple Storage Service, known as Amazon S3, and the Amazon Elastic Compute Cloud (EC2) were released.[8][9]

    The following decade saw the launch of various cloud services. In 2010, Microsoft launched Microsoft Azure, and Rackspace Hosting and NASA initiated an open-source cloud-software project, OpenStack. IBM introduced the IBM SmartCloud framework in 2011, and Oracle announced the Oracle Cloud in 2012. In December 2019, Amazon launched AWS Outposts, a service that extends AWS infrastructure, services, APIs, and tools to customer data centers, co-location spaces, or on-premises facilities.[10][11]

    Value proposition

    This section cites its sources but does not provide page references. Please help improve it by providing page numbers for existing citations. (January 2024) (Learn how and when to remove this message)
    Cloud computing can shorten time to market by offering pre-configured tools, scalable resources, and managed services, allowing users to focus on core business value rather than maintaining infrastructure. Cloud platforms can enable organizations and individuals to reduce upfront capital expenditures on physical infrastructure by shifting to an operational expenditure model, where costs scale with usage. Cloud platforms also offer managed services and tools, such as artificial intelligence, data analytics, and machine learning, which might otherwise require significant in-house expertise and infrastructure investment.[12][13][14]

    While cloud computing can offer cost advantages through effective resource optimization, organizations often face challenges such as unused resources, inefficient configurations, and hidden costs without proper oversight and governance. Many cloud platforms provide cost management tools, such as AWS Cost Explorer and Azure Cost Management, and frameworks like FinOps have emerged to standardize financial operations in the cloud. Cloud computing also facilitates collaboration, remote work, and global service delivery by enabling secure access to data and applications from any location with an internet connection.[12][13][14]

    Cloud providers offer various redundancy options for core services, such as managed storage and managed databases, though redundancy configurations often vary by service tier. Advanced redundancy strategies, such as cross-region replication or failover systems, typically require explicit configuration and may incur additional costs or licensing fees.[12][13][14]

    Cloud environments operate under a shared responsibility model, where providers are typically responsible for infrastructure security, physical hardware, and software updates, while customers are accountable for data encryption, identity and access management (IAM), and application-level security. These responsibilities vary depending on the cloud service model—Infrastructure as a Service (IaaS), Platform as a Service (PaaS), or Software as a Service (SaaS)—with customers typically having more control and responsibility in IaaS environments and progressively less in PaaS and SaaS models, often trading control for convenience and managed services.[12][13][14]

    Adoption and suitability

    This section cites its sources but does not provide page references. Please help improve it by providing page numbers for existing citations. (January 2024) (Learn how and when to remove this message)
    The decision to adopt cloud computing or maintain on-premises infrastructure depends on factors such as scalability, cost structure, latency requirements, regulatory constraints, and infrastructure customization.[15][16][17][18]

    Organizations with variable or unpredictable workloads, limited capital for upfront investments, or a focus on rapid scalability benefit from cloud adoption. Startups, SaaS companies, and e-commerce platforms often prefer the pay-as-you-go operational expenditure (OpEx) model of cloud infrastructure. Additionally, companies prioritizing global accessibility, remote workforce enablement, disaster recovery, and leveraging advanced services such as AI/ML and analytics are well-suited for the cloud. In recent years, some cloud providers have started offering specialized services for high-performance computing and low-latency applications, addressing some use cases previously exclusive to on-premises setups.[15][16][17][18]

    On the other hand, organizations with strict regulatory requirements, highly predictable workloads, or reliance on deeply integrated legacy systems may find cloud infrastructure less suitable. Businesses in industries like defense, government, or those handling highly sensitive data often favor on-premises setups for greater control and data sovereignty. Additionally, companies with ultra-low latency requirements, such as high-frequency trading (HFT) firms, rely on custom hardware (e.g., FPGAs) and physical proximity to exchanges, which most cloud providers cannot fully replicate despite recent advancements. Similarly, tech giants like Google, Meta, and Amazon build their own data centers due to economies of scale, predictable workloads, and the ability to customize hardware and network infrastructure for optimal efficiency. However, these companies also use cloud services selectively for certain workloads and applications where it aligns with their operational needs.[15][16][17][18]

    In practice, many organizations are increasingly adopting hybrid cloud architectures, combining on-premises infrastructure with cloud services. This approach allows businesses to balance scalability, cost-effectiveness, and control, offering the benefits of both deployment models while mitigating their respective limitations.[15][16][17][18]

    Challenges and limitations
    Main article: Cloud computing issues
    One of the primary challenges of cloud computing, compared with traditional on-premises systems, is maintaining data security and privacy. Cloud users entrust their sensitive data to third-party providers, who may not have adequate measures to protect it from unauthorized access, breaches, or leaks. Cloud users also face compliance risks if they have to adhere to certain regulations or standards regarding data protection, such as GDPR or HIPAA.[19]

    Another challenge of cloud computing is reduced visibility and control. Cloud users may not have full insight into how their cloud resources are managed, configured, or optimized by their providers. They may also have limited ability to customize or modify their cloud services according to their specific needs or preferences.[19] Complete understanding of all technology may be impossible, especially given the scale, complexity, and deliberate opacity of contemporary systems; however, there is a need for understanding complex technologies and their interconnections to have power and agency within them.[20] The metaphor of the cloud can be seen as problematic as cloud computing retains the aura of something noumenal and numinous; it is something experienced without precisely understanding what it is or how it works.[21]

    Additionally, cloud migration is a significant challenge. This process involves transferring data, applications, or workloads from one cloud environment to another, or from on-premises infrastructure to the cloud. Cloud migration can be complicated, time-consuming, and expensive, particularly when there are compatibility issues between different cloud platforms or architectures. If not carefully planned and executed, cloud migration can lead to downtime, reduced performance, or even data loss.[22]

    Cloud migration challenges
    According to the 2024 State of the Cloud Report by Flexera, approximately 50% of respondents identified the following top challenges when migrating workloads to public clouds:[23]

    “Understanding application dependencies”
    “Comparing on-premise and cloud costs”
    “Assessing technical feasibility.”
    Implementation challenges
    Applications hosted in the cloud are susceptible to the fallacies of distributed computing, a series of misconceptions that can lead to significant issues in software development and deployment.[24]

    Cloud cost overruns
    In a report by Gartner, a survey of 200 IT leaders revealed that 69% experienced budget overruns in their organizations’ cloud expenditures during 2023. Conversely, 31% of IT leaders whose organizations stayed within budget attributed their success to accurate forecasting and budgeting, proactive monitoring of spending, and effective optimization.[25]

    The 2024 Flexera State of Cloud Report identifies the top cloud challenges as managing cloud spend, followed by security concerns and lack of expertise. Public cloud expenditures exceeded budgeted amounts by an average of 15%. The report also reveals that cost savings is the top cloud initiative for 60% of respondents. Furthermore, 65% measure cloud progress through cost savings, while 42% prioritize shorter time-to-market, indicating that cloud’s promise of accelerated deployment is often overshadowed by cost concerns.[23]

    Service Level Agreements
    Typically, cloud providers’ Service Level Agreements (SLAs) do not encompass all forms of service interruptions. Exclusions typically include planned maintenance, downtime resulting from external factors such as network issues, human errors, like misconfigurations, natural disasters, force majeure events, or security breaches. Typically, customers bear the responsibility of monitoring SLA compliance and must file claims for any unmet SLAs within a designated timeframe. Customers should be aware of how deviations from SLAs are calculated, as these parameters may vary by service. These requirements can place a considerable burden on customers. Additionally, SLA percentages and conditions can differ across various services within the same provider, with some services lacking any SLA altogether. In cases of service interruptions due to hardware failures in the cloud provider, the company typically does not offer monetary compensation. Instead, eligible users may receive credits as outlined in the corresponding SLA.[26][27][28][29]

    Leaky abstractions
    Cloud computing abstractions aim to simplify resource management, but leaky abstractions can expose underlying complexities. These variations in abstraction quality depend on the cloud vendor, service and architecture. Mitigating leaky abstractions requires users to understand the implementation details and limitations of the cloud services they utilize.[30][31][32]

    Service lock-in within the same vendor
    Service lock-in within the same vendor occurs when a customer becomes dependent on specific services within a cloud vendor, making it challenging to switch to alternative services within the same vendor when their needs change.[33][34]

    Security and privacy

    Cloud suppliers security and privacy agreements must be aligned to the demand(s) requirements and regulations.
    Main article: Cloud computing security
    Cloud computing poses privacy concerns because the service provider can access the data that is in the cloud at any time. It could accidentally or deliberately alter or delete information.[35] Many cloud providers can share information with third parties if necessary for purposes of law and order without a warrant. That is permitted in their privacy policies, which users must agree to before they start using cloud services. Solutions to privacy include policy and legislation as well as end-users’ choices for how data is stored.[35] Users can encrypt data that is processed or stored within the cloud to prevent unauthorized access.[35] Identity management systems can also provide practical solutions to privacy concerns in cloud computing. These systems distinguish between authorized and unauthorized users and determine the amount of data that is accessible to each entity.[36] The systems work by creating and describing identities, recording activities, and getting rid of unused identities.

    According to the Cloud Security Alliance, the top three threats in the cloud are Insecure Interfaces and APIs, Data Loss & Leakage, and Hardware Failure—which accounted for 29%, 25% and 10% of all cloud security outages respectively. Together, these form shared technology vulnerabilities. In a cloud provider platform being shared by different users, there may be a possibility that information belonging to different customers resides on the same data server. Additionally, Eugene Schultz, chief technology officer at Emagined Security, said that hackers are spending substantial time and effort looking for ways to penetrate the cloud. “There are some real Achilles’ heels in the cloud infrastructure that are making big holes for the bad guys to get into”. Because data from hundreds or thousands of companies can be stored on large cloud servers, hackers can theoretically gain control of huge stores of information through a single attack—a process he called “hyperjacking”. Some examples of this include the Dropbox security breach, and iCloud 2014 leak.[37] Dropbox had been breached in October 2014, having over seven million of its users passwords stolen by hackers in an effort to get monetary value from it by Bitcoins (BTC). By having these passwords, they are able to read private data as well as have this data be indexed by search engines (making the information public).[37]

    There is the problem of legal ownership of the data (If a user stores some data in the cloud, can the cloud provider profit from it?). Many Terms of Service agreements are silent on the question of ownership.[38] Physical control of the computer equipment (private cloud) is more secure than having the equipment off-site and under someone else’s control (public cloud). This delivers great incentive to public cloud computing service providers to prioritize building and maintaining strong management of secure services.[39] Some small businesses that do not have expertise in IT security could find that it is more secure for them to use a public cloud. There is the risk that end users do not understand the issues involved when signing on to a cloud service (persons sometimes do not read the many pages of the terms of service agreement, and just click “Accept” without reading). This is important now that cloud computing is common and required for some services to work, for example for an intelligent personal assistant (Apple’s Siri or Google Assistant). Fundamentally, private cloud is seen as more secure with higher levels of control for the owner, however public cloud is seen to be more flexible and requires less time and money investment from the user.[40]

    The attacks that can be made on cloud computing systems include man-in-the middle attacks, phishing attacks, authentication attacks, and malware attacks. One of the largest threats is considered to be malware attacks, such as Trojan horses. Recent research conducted in 2022 has revealed that the Trojan horse injection method is a serious problem with harmful impacts on cloud computing systems.[41]

    Extraterritorial data access
    Main article: Data sovereignty § United States extraterritorial data access laws
    The CLOUD Act allows United States authorities to request data from cloud providers, and courts can impose nondisclosure requirements preventing providers from notifying affected users.[42] This framework is in legal tension with Article 48 of the European General Data Protection Regulation (GDPR), which restricts the transfer of personal data in response to foreign court or administrative orders unless based on an international agreement. As a result, cloud service providers operating in both Europe and the U.S. may face competing legal obligations.[43]

    According to Laura K. Donohue writing for the Harvard Journal of Law and Public Policy, cloud service providers also fall within the broader category of service providers subject to Section 702 of the Foreign Intelligence Surveillance Act (FISA), which has had documented effects on cloud providers and their customers.[44]

    Service models
    See also: “as a service”

    Comparison of on-premise, IaaS, PaaS, and SaaS

    Cloud computing service models arranged as layers in a stack
    The National Institute of Standards and Technology recognized three cloud service models in 2011: Infrastructure as a Service (IaaS), Platform as a Service (PaaS), and Software as a Service (SaaS).[3] The International Organization for Standardization (ISO) later identified additional models in 2023, including “Network as a Service”, “Communications as a Service”, “Compute as a Service”, and “Data Storage as a Service”.[4]

    Infrastructure as a service (IaaS)
    Main article: Infrastructure as a service
    Infrastructure as a service (IaaS) refers to online services that provide high-level APIs used to abstract various low-level details of underlying network infrastructure like physical computing resources, location, data partitioning, scaling, security, backup, etc. A hypervisor runs the virtual machines as guests. Pools of hypervisors within the cloud operational system can support large numbers of virtual machines and the ability to scale services up and down according to customers’ varying requirements. Linux containers run in isolated partitions of a single Linux kernel running directly on the physical hardware. Linux cgroups and namespaces are the underlying Linux kernel technologies used to isolate, secure and manage the containers. The use of containers offers higher performance than virtualization because there is no hypervisor overhead. IaaS clouds often offer additional resources such as a virtual-machine disk-image library, raw block storage, file or object storage, firewalls, load balancers, IP addresses, virtual local area networks (VLANs), and software bundles.[45]

    Further information: CloudBioLinux
    The NIST’s definition of cloud computing describes IaaS as “where the consumer is able to deploy and run arbitrary software, which can include operating systems and applications. The consumer does not manage or control the underlying cloud infrastructure but has control over operating systems, storage, and deployed applications; and possibly limited control of select networking components (e.g., host firewalls).”[3]

    IaaS-cloud providers supply these resources on-demand from their large pools of equipment installed in data centers. For wide-area connectivity, customers can use either the Internet or carrier clouds (dedicated virtual private networks). To deploy their applications, cloud users install operating-system images and their application software on the cloud infrastructure. In this model, the cloud user patches and maintains the operating systems and the application software. Cloud providers typically bill IaaS services on a utility computing basis: cost reflects the number of resources allocated and consumed.[46]

    Platform as a service (PaaS)
    Main article: Platform as a service
    The NIST’s definition of cloud computing defines Platform as a Service as:[3]

    The capability provided to the consumer is to deploy onto the cloud infrastructure consumer-created or acquired applications created using programming languages, libraries, services, and tools supported by the provider. The consumer does not manage or control the underlying cloud infrastructure including network, servers, operating systems, or storage, but has control over the deployed applications and possibly configuration settings for the application-hosting environment.

    PaaS vendors offer a development environment to application developers. The provider typically develops toolkit and standards for development and channels for distribution and payment. In the PaaS models, cloud providers deliver a computing platform, typically including an operating system, programming-language execution environment, database, and the web server. Application developers develop and run their software on a cloud platform instead of directly buying and managing the underlying hardware and software layers. With some PaaS, the underlying computer and storage resources scale automatically to match application demand so that the cloud user does not have to allocate resources manually.[47][need quotation to verify]

    Some integration and data management providers also use specialized applications of PaaS as delivery models for data. Examples include iPaaS (Integration Platform as a Service) and dPaaS (Data Platform as a Service). iPaaS enables customers to develop, execute and govern integration flows.[48] Under the iPaaS integration model, customers drive the development and deployment of integrations without installing or managing any hardware or middleware.[49] dPaaS delivers integration—and data-management—products as a fully managed service.[50] Under the dPaaS model, the PaaS provider, not the customer, manages the development and execution of programs by building data applications for the customer. dPaaS users access data through data-visualization tools.[51]

    Software as a service (SaaS)
    Main article: Software as a service
    The NIST’s definition of cloud computing defines Software as a Service as:[3]

    The capability provided to the consumer is to use the provider’s applications running on a cloud infrastructure. The applications are accessible from various client devices through either a thin client interface, such as a web browser (e.g., web-based email), or a program interface. The consumer does not manage or control the underlying cloud infrastructure including network, servers, operating systems, storage, or even individual application capabilities, with the possible exception of limited user-specific application configuration settings.

    In the software as a service (SaaS) model, users gain access to application software and databases. Cloud providers manage the infrastructure and platforms that run the applications. SaaS is sometimes referred to as “on-demand software” and is usually priced on a pay-per-use basis or using a subscription fee.[52] In the SaaS model, cloud providers install and operate application software in the cloud and cloud users access the software from cloud clients. Cloud users do not manage the cloud infrastructure and platform where the application runs. This eliminates the need to install and run the application on the cloud user’s own computers, which simplifies maintenance and support. Cloud applications differ from other applications in their scalability—which can be achieved by cloning tasks onto multiple virtual machines at run-time to meet changing work demand.[53] Load balancers distribute the work over the set of virtual machines. This process is transparent to the cloud user, who sees only a single access-point. To accommodate a large number of cloud users, cloud applications can be multitenant, meaning that any machine may serve more than one cloud-user organization.

    The pricing model for SaaS applications is typically a monthly or yearly flat fee per user,[54] so prices become scalable and adjustable if users are added or removed at any point. It may also be free.[55] Proponents claim that SaaS gives a business the potential to reduce IT operational costs by outsourcing hardware and software maintenance and support to the cloud provider. This enables the business to reallocate IT operations costs away from hardware/software spending and from personnel expenses, towards meeting other goals. In addition, with applications hosted centrally, updates can be released without the need for users to install new software. One drawback of SaaS comes with storing the users’ data on the cloud provider’s server. As a result,[citation needed] there could be unauthorized access to the data.[56] Examples of applications offered as SaaS are games and productivity software like Google Docs and Office Online. SaaS applications may be integrated with cloud storage or File hosting services, which is the case with Google Docs being integrated with Google Drive, and Office Online being integrated with OneDrive.[57]

    Serverless computing
    Main article: Serverless computing
    Serverless computing allows customers to use various cloud capabilities without the need to provision, deploy, or manage hardware or software resources, apart from providing their application code or data. ISO/IEC 22123-2:2023 classifies serverless alongside Infrastructure as a Service (IaaS), Platform as a Service (PaaS), and Software as a Service (SaaS) under the broader category of cloud service categories. Notably, while ISO refers to these classifications as cloud service categories, the National Institute of Standards and Technology (NIST) refers to them as service models.[3][4]

    Deployment models

    Cloud computing types
    “A cloud deployment model represents the way in which cloud computing can be organized based on the control and sharing of physical or virtual resources.”[4] Cloud deployment models define the fundamental patterns of interaction between cloud customers and cloud providers. They do not detail implementation specifics or the configuration of resources.[4]

    Private
    Private cloud is cloud infrastructure operated solely for a single organization, whether managed internally or by a third party, and hosted either internally or externally.[3] Undertaking a private cloud project requires significant engagement to virtualize the business environment, and requires the organization to reevaluate decisions about existing resources. It can improve business, but every step in the project raises security issues that must be addressed to prevent serious vulnerabilities. Self-run data centers[58] are generally capital intensive. They have a significant physical footprint, requiring allocations of space, hardware, and environmental controls. These assets have to be refreshed periodically, resulting in additional capital expenditures. They have attracted criticism because users “still have to buy, build, and manage them” and thus do not benefit from less hands-on management,[59] essentially “[lacking] the economic model that makes cloud computing such an intriguing concept”.[60][61]

    Public
    For a comparison of cloud-computing software and providers, see Cloud-computing comparison
    Cloud services are considered “public” when they are delivered over the public Internet, and they may be offered as a paid subscription, or free of charge.[62] Architecturally, there are few differences between public- and private-cloud services, but security concerns increase substantially when services (applications, storage, and other resources) are shared by multiple customers. Most public-cloud providers offer direct-connection services that allow customers to securely link their legacy data centers to their cloud-resident applications.[63][64]

    Several factors like the functionality of the solutions, cost, integrational and organizational aspects as well as safety & security are influencing the decision of enterprises and organizations to choose a public cloud or on-premises solution.[65]

    Hybrid
    See also: Hybrid cloud storage
    Hybrid cloud is a composition of a public cloud and a private environment, such as a private cloud or on-premises resources,[66][67] that remain distinct entities but are bound together, offering the benefits of multiple deployment models. Hybrid cloud can also mean the ability to connect collocation, managed or dedicated services with cloud resources.[3] Gartner defines a hybrid cloud service as a cloud computing service that is composed of some combination of private, public and community cloud services, from different service providers.[68] A hybrid cloud service crosses isolation and provider boundaries so that it cannot be simply put in one category of private, public, or community cloud service. It allows one to extend either the capacity or the capability of a cloud service, by aggregation, integration or customization with another cloud service.

    Varied use cases for hybrid cloud composition exist. For example, an organization may store sensitive client data in house on a private cloud application, but interconnect that application to a business intelligence application provided on a public cloud as a software service.[69] This example of hybrid cloud extends the capabilities of the enterprise to deliver a specific business service through the addition of externally available public cloud services. Hybrid cloud adoption depends on a number of factors such as data security and compliance requirements, level of control needed over data, and the applications an organization uses.[70]

    Another example of hybrid cloud is one where IT organizations use public cloud computing resources to meet temporary capacity needs that can not be met by the private cloud.[71] This capability enables hybrid clouds to employ cloud bursting for scaling across clouds.[3] Cloud bursting is an application deployment model in which an application runs in a private cloud or data center and “bursts” to a public cloud when the demand for computing capacity increases. A primary advantage of cloud bursting and a hybrid cloud model is that an organization pays for extra compute resources only when they are needed.[72] Cloud bursting enables data centers to create an in-house IT infrastructure that supports average workloads, and use cloud resources from public or private clouds, during spikes in processing demands.[73]

    Community
    Community cloud shares infrastructure between several organizations from a specific community with common concerns (security, compliance, jurisdiction, etc.), whether it is managed internally or by a third-party, and hosted internally or externally, the costs are distributed among fewer users compared to a public cloud (but more than a private cloud). As a result, only a portion of the potential cost savings of cloud computing is achieved. [3]

    Multi cloud
    Main article: Multicloud
    According to ISO/IEC 22123-1: “multi-cloud is a cloud deployment model in which a customer uses public cloud services provided by two or more cloud service providers”. [74] Poly cloud refers to the use of multiple public clouds for the purpose of leveraging specific services that each provider offers. It differs from Multi cloud in that it is not designed to increase flexibility or mitigate against failures but is rather used to allow an organization to achieve more than could be done with a single provider.[75]

    Market
    According to recent market research estimates, the global cloud computing market was approximately US$1.0–1.2 trillion in 2026.[76] According to International Data Corporation, global public cloud services spending is forecast to surpass US$1 trillion in 2026.[77] According to Gartner, worldwide public cloud services spending is forecast to reach US$1.03 trillion in 2026.[78]

    Market research forecasts indicate that the global cloud computing market could grow substantially over the following decade, with estimates for 2035 ranging from approximately US$3.59 trillion to US$6 trillion.[79]

    The European Commission’s 2012 Communication identified several issues which were impeding the development of the cloud computing market:[80]: Section 3 

    fragmentation of the digital single market across the EU
    concerns about contracts including reservations about data access and ownership, data portability, and change control
    variations in standards applicable to cloud computing
    The Communication set out a series of “digital agenda actions” which the Commission proposed to undertake in order to support the development of a fair and effective market for cloud computing services.[80]: Pages 6–14 

    Similar concepts
    The goal of cloud computing is to allow users to take benefit from all of these technologies, without the need for deep knowledge about or expertise with each one of them. The cloud aims to cut costs and helps the users focus on their core business instead of being impeded by IT obstacles.[81] The main enabling technology for cloud computing is virtualization. Virtualization software separates a physical computing device into one or more “virtual” devices, each of which can be easily used and managed to perform computing tasks. With operating system-level virtualization essentially creating a scalable system of multiple independent computing devices, idle computing resources can be allocated and used more efficiently. Virtualization provides the agility required to speed up IT operations and reduces cost by increasing infrastructure utilization. Autonomic computing automates the process through which the user can provision resources on-demand. By minimizing user involvement, automation speeds up the process, reduces labor costs and reduces the possibility of human errors.[81]

    Cloud computing uses concepts from utility computing to provide metrics for the services used. Cloud computing attempts to address QoS (quality of service) and reliability problems of other grid computing models.[81]

    Cloud computing shares characteristics with:

    Client–server model – Client–server computing refers broadly to any distributed application that distinguishes between service providers (servers) and service requestors (clients).[82]
    Computer bureau – A service bureau providing computer services, particularly from the 1960s to 1980s.
    Grid computing – A form of distributed and parallel computing, whereby a ‘super and virtual computer’ is composed of a cluster of networked, loosely coupled computers acting in concert to perform very large tasks.
    Fog computing – Distributed computing paradigm that provides data, compute, storage and application services closer to the client or near-user edge devices, such as network routers. Furthermore, fog computing handles data at the network level, on smart devices and on the end-user client-side (e.g. mobile devices), instead of sending data to a remote location for processing.
    Utility computing – The “packaging of computing resources, such as computation and storage, as a metered service similar to a traditional public utility, such as electricity.”[83][84]
    Peer-to-peer – A distributed architecture without the need for central coordination. Participants are both suppliers and consumers of resources (in contrast to the traditional client-server model).
    Cloud sandbox – A live, isolated computer environment in which a program, code or file can run without affecting the application in which it runs.

  • Robotics

    Robotics is the interdisciplinary study and practice of the design, construction, operation, and use of robots.[1] A roboticist is someone who specializes in robotics.[2] Robotics usually combines four aspects of design work: a power source (e.g. a battery), mechanical construction, a control system (electrical circuits), and software (run by remote control or artificial intelligence).

    The goal of most robotics is to design machines that can assist humans in various fields, such as agriculture, construction, domestic work, food processing, inventory management, manufacturing, medicine, military, mining, space exploration, and transportation.

    Robots impact humans by displacing workers. Some expect this to occur at an increasing rate, leading to proposed solutions such as basic income. Robotics is itself a lucrative business that creates careers, especially for postgraduates. Roboticists often aim to create machines that seem to interface naturally with humans. The field is under active research and development, with areas of interest including robot kinematics and quantum robotics.

    Design

    Robotics usually combines four aspects of design work to create a robot:

    • Power source: Potential energy sources include wired electricity, a battery, and/or petrol.
    • Mechanical construction: A physical form or combination of forms is designed to functionally achieve tasks within a given range of environments. This can include locomotive elements such as wheels and caterpillar tracks, as well as hydraulic limbs and manipulators (e.g. hands).
    • Control system: Electrical circuits (utilizing components such as diodes and transistors) are used to run software, govern motor movement, and read sensors.
    • Software: A program is how a robot decides when or how to do something. Robotic programs can be run by remote control, artificial intelligence (AI), or a hybrid of the two. AI programming is an important part of robotic navigation and human–robot interaction.

    Power source

    Further information: Power supply and Energy storage

    A mobile robot is ideally powered by lightweight batteries with a high lifetime of recharge cycles, with safety considered both related to active operation and the effects of waste on the environment.[3] Various kinds of batteries can be used, with lead–acid batteries being the most well-established, safe, and relatively long-lived, although they are relatively cumbersome. Silver–cadmium batteries can be much smaller but have limited recharges and are quite expensive; they are thus mostly used for space and military operations.[3] A battery-replacement system prevents downtime during charge cycles.[4]

    Generators, often some type of internal combustion engine, can also be used, but are often mechanically complex and inefficient.[3] A tether can connect a robot to its power supply, saving weight and space but likely stymied by cable-spooling difficulties.[3] Solar power is most effective for space or stationary robots due to a slow charge cycle.[3] Other potential power sources include:

    Mechanical construction

    See also: Mechanical engineering

    A robotic leg powered by air muscles

    Actuators convert stored energy into movement.[5] They are generally limited to the location of a joint, as compared to the muscles of a biological organism, forcing robot designers to be creative to maximize power without breaking parts.[6]

    The most popular actuators are electric motors that rotate a wheel or gear[7] and linear actuators that control factory robots. Most robots use electric motors, often brushed and brushless DC motors in portable robots[6] or AC motors in industrial robots and computer numerical control machines (especially in systems with lighter loads and predominantly rotational motion). Meanwhile, linear actuators simply move in and out (being used in industrial applications). They are typically powered by oil or compressed air,[8][9] but can also be powered by electricity, usually via a motor and a leadscrew.[10] The mechanical rack and pinion is common in large gantry robots.[11]

    Recent alternatives to DC motors are piezoelectric motors, including ultrasonic motors, in which tiny piezoceramic elements vibrate many thousands of times per second, causing linear or rotary motion. One type uses the vibration of the piezo elements to step the motor in a circle or a straight line;[12] another type uses the piezo elements to vibrate a nut or drive a screw. The advantages of these motors are nanometer resolution, speed, and force for their size.[13][14][15]

    Series elastic actuation (SEA) relies on introducing intentional elasticity between the motor actuator and the load for robust force control. Due to the resultant lower reflected inertia, series elastic actuation improves safety during robot interactions or collisions.[16] Further, it provides energy efficiency and shock absorption (mechanical filtering) while reducing excessive wear on the transmission and other components. This approach has successfully been employed in various robots, particularly advanced manufacturing robots[17] and walking humanoid robots.[18][19] The controller design of a series elastic actuator is most often performed within the passivity framework as it ensures the safety of interaction with unstructured environments.[20]

    Pneumatic artificial muscles, also known as air muscles, are special tubes that expand (typically up to 42%) when air is forced inside them; they are used in some robot applications.[21][22][23] Muscle wire, also known as shape memory alloy, is a material that contracts (under 5%) when electricity is applied; they have been used for some small robots.[24][25] Electroactive polymers are a plastic material that can contract substantially (up to 380% activation strain) from electricity and have been used in the facial muscles and arms of humanoid robots,[26] as well as to enable new robots to float,[27] fly, swim or walk.[28] Additionally, elastic carbon nanotubes are a promising experimental artificial muscle technology. The absence of defects in carbon nanotubes enables these filaments to deform elastically by several percent, with energy storage levels of perhaps 10 J/cm3 for metal nanotubes. Human biceps could be replaced with wire of this material measuring 8 millimetres (3⁄8 in) in diameter, feasibly allowing future robots to outperform humans.[29]

    Locomotion

    Main articles: Robot locomotion and Mobile robot

    Robots with only one or two wheel(s) can have advantages such as greater efficiency, reduced parts, and navigation through confined areas. A one-wheeled robot balances on a round ball; Carnegie Mellon University‘s Ballbot is the approximate height and width of a person.[30][31] Several attempts have also been made to build spherical robots (also known as orb bots[32] or ball bots),[33] which move by spinning a weight inside the ball[34][35] or rotating outer shells.[36][37] Two-wheeled balancing robots generally use a gyroscope to detect how much a robot is falling and drive the wheels proportionally up to hundreds of times per second to counterbalance the fall, based on inverted pendulum dynamics.[38][39] NASA‘s Robonaut has been mounted to a Segway for a similar effect.[40] Most mobile robots have four wheels or continuous tracks. Six wheels can give better traction in outdoor terrain, while tracks provide even more grip. Tracked wheels are common for outdoor off-road robots, but are difficult to use indoors.[41] A small number of skating robots have been developed, one of which is a multimodal walking and skating device with four legs and unpowered wheels.[42][43]

    Mantis the spider robot in 2012

    Several robots have been made that can walk on two legs, but not yet as reliably as a human.[44] Many other robots have been built that walk on more than two legs, being significantly easier.[45][46] Walking robots could be used for uneven terrains, providing a high degree of mobility and efficiency, but two-legged robots can currently only handle flat floors or perhaps stairs. Some approaches have included:

    • The zero moment point (ZMP) is the algorithm used by robots such as Honda‘s ASIMO. The robot’s onboard computer tries to keep the total inertial forces (the combination of Earth’s gravity and the acceleration and deceleration of walking) exactly opposed by the floor reaction force (the force of the floor pushing back on the robot’s foot). In this way, the two forces cancel out, leaving no moment (force causing the robot to rotate and fall over).[47] Human observers note that this is not exactly how a human walks, with some describing ASIMO’s walk as looking like it needs use the bathroom.[48][49][50] ASIMO’s walking algorithm utilizes some dynamic balancing, but requires a flat surface.
    • Several robots, built in the 1980s by Marc Raibert at the MIT Leg Laboratory, successfully demonstrated very dynamic walking. Initially, a robot with only one leg, and a very small foot could stay upright simply by hopping. The movement is the same as that of a person on a pogo stick. As the robot falls to one side, it would jump slightly in that direction to catch itself.[51] Soon, the algorithm was generalized to two and four legs. A bipedal robot was demonstrated running and even performing somersaults.[52] A quadruped was also demonstrated which could trot, run, pace, and bound.[53][54]
    • A more advanced approach is a dynamic balancing algorithm, which constantly monitors the robot’s motion and places the feet to maintain stability.[55] This technique has been demonstrated by Anybots‘ Dexter robot[56] (which is so stable it can perform jumps)[57] and Delft University‘s Flame.
    • Perhaps the most promising approach uses passive dynamics, in which the momentum of swinging limbs is used to power walking, perhaps increasing efficiency to ten times that of ZMP.[58][59]
    Visualization of entomopter flying on Mars (NASA)

    A modern passenger airliner is essentially a flying robot, with two humans to manage it. The autopilot can control the plane through takeoff, normal flight, and landing.[60] Unmanned aerial vehicles (UAVs) can be smaller and lighter and fly into dangerous territory for military use, perhaps even being triggered to fire automatically. Other flying robots include cruise missiles, the entomopter, and the Epson micro helicopter robot. Additionally, some lighter-than-air robots are propelled by paddles and guided by sonar.

    Biomimetic flying robots (BFRs) take inspiration from flying mammals, birds, or insects. They can have flapping wings, which generate the lift and thrust, or they can be propeller-actuated. Flapping-wing designs have increased maneuverability and reduced energy consumption compared to propeller actuation.[61] BFRs inspired by mammals and birds share similar flight characteristics and design considerations. For instance, they minimize edge fluttering and pressure-induced wingtip curl by increasing the rigidity of the wing edge.

    • Mammal-inspired BFRs typically take inspiration from bats, with the flying squirrel also inspiring a prototype.[62][63][64] Mammal-inspired BFRs can be designed to be multimodal; being capable of both flight and terrestrial movement. Shock absorbers can be implemented to reduce the impact of landing.[64] Alternatively, the BFR can pitch up and increase the amount of drag.[62] Different land gait patterns can also be implemented.[62]
    • Bird-inspired BFRs can take inspiration from raptors, gulls, and others.[65][66] They can be feathered to increase the angle of range over which the robot can operate before stalling.[67] The wings of bird-inspired BFRs allow for in-plane deformation, which can be adjusted to maximize flight efficiency depending on the flight gait.[67]
    • Insect-inspired BFRs typically take inspiration from beetles or dragonflies.[68][69][70]
    Capuchin, a climbing robot

    Several different approaches have been used to develop robots that have the ability to climb vertical surfaces. One approach mimics the movements of a human climber on a wall with protrusions; adjusting the center of mass and moving each limb in turn to gain leverage.[71] Another approach uses the specialized toe-pad method of wall-climbing geckoes, which can run on smooth surfaces such as vertical glass,[72][73] one example being named Speedy Freelander. A third approach is to mimic the motion of a snake climbing a pole.[74] Separately, snake robots can be used for horizontal navigation, possibly being able to search through confined spaces[75] and navigate amphibiously.[76][77]

    It is calculated that when swimming, some fish can achieve a propulsive efficiency greater than 90%.[78] Furthermore, they can accelerate and maneuver far better than any man-made boat or submarine and cause less disturbance, being a desirable ability for aquatic robots,[79][80] one of which models fish locomotion.[81] One example copies the streamlined shape and propulsion of the front ‘flippers’ of penguins.[82] Others emulate the locomotion of the manta ray and jellyfish. In 2013, a robotic fish outperformed some real fish in average maximum velocity and endurance.[83][84][85]

    Sailboat robots, such as Vaimos, have been developed in order to make measurements at the surface of the ocean.[86] Since saiboat robots are wind-propelled, the batteries only power the computer, communication and actuators (to tune the rudder and sail). Two major sailboat robot competitions occur at the Microtransat Challenge and the World Robotic Sailing Championship.

    Manipulators

    Further information: Mobile manipulator

    Baxter, a robot with versatile arms
    A robotic hand

    A definition of robotic manipulation has been described by Matthew T. Mason as the robot’s “control of its environment through selective contact”.[87] Robots need to manipulate objects; pick up, modify, destroy, move or otherwise have an effect. Thus a robotic arm is referred to as a manipulator[88] and its functional end (e.g. a tool or hand) is known as an end effector.[89] Most robot arms have replaceable end effectors, each allowing them to perform some small range of tasks. Some have a fixed manipulator that cannot be replaced, including highly versatile manipulators like a humanoid hand.[90][91][92] Some of these have powerful dexterity intelligence, up to 20 degrees of freedom, and hundreds of tactile sensors.[93]

    One of the most common types of end effectors are ‘grippers’. In its simplest manifestation, it consists of just two fingers that can open and close to pick up and let go of small objects. Fingers can be made of a chain with a metal wire running through it.[94] Hands that resemble and work more like a human hand include the Shadow Hand and the Robonaut hand.[95][96][97] Mechanical grippers can come in various types, including friction and encompassing jaws. Friction jaws use all the force of the gripper to hold the object in place using friction. Encompassing jaws cradle the object in place, using less friction.

    Suction end effectors, powered by vacuum generators, are very simple astrictive[98] devices that can hold very large loads provided the prehension surface is smooth enough to ensure suction. Pick-and-place robots for electronic components and for large objects like car windscreens, often use very simple vacuum end effectors. Suction is a highly used type of end effector in industry, in part because the natural compliance of soft suction end effectors can be less likely to damage objects.

    Control system

    Further information: Control system and Principles of motion sensing

    An electrical circuit

    The mechanical structure of a robot must be controlled to perform tasks.[99] The control of a robot involves three distinct phases – perception, processing, and action (robotic paradigms).[100] Sensors give information about the environment or the robot itself (e.g. the position of its joints or its end effector).[101] This information is then processed to be stored or transmitted and to calculate the appropriate signals to the actuators (motors), which move the mechanical structure to achieve the required coordinated motion or force actions.

    The processing phase can range in complexity. At a reactive level, it may translate raw sensor information directly into actuator commands (e.g. firing motor power electronic gates based directly upon encoder feedback signals to achieve the required torque/velocity of the shaft). Sensor fusion and internal models may first be used to estimate parameters of interest (e.g. the position of the robot’s gripper) from noisy sensor data. An immediate task (such as moving the gripper in a certain direction until an object is detected with a proximity sensor) is sometimes inferred from these estimates. Techniques from control theory are generally used to convert the higher-level tasks into individual commands that drive the actuators, most often using kinematic and dynamic models of the mechanical structure.[99][100][102]

    At longer time scales or with more sophisticated tasks, the robot may need to build and reason with a cognitive model, which try to represent the robot, the world, and how the two interact. Pattern recognition and computer vision can be used to track objects.[99] Mapping, motion planning and other AI techniques may be used to figure out how to act and avoid obstacles.

    Robotic control systems integrate multiple sensors and effectors, have many interacting degrees of freedom and require operator interfaces, programming tools and real-time capabilities.[100] They are often connected to wider communication networks, including the Internet of things, a network correlating physical objects.[103] Progress towards open architecture, layered, user-friendly and ‘intelligent’ sensor-based interconnected robots has emerged from earlier concepts related to flexible manufacturing systems. Further, several ‘open or ‘hybrid’ reference architectures provide advantages over prior ‘closed’ robot control systems.[102] Open-architecture controllers are said to be better able to meet the growing requirements of a wide range of robot users, including system developers, end users and research scientists, and are better positioned to contribute advanced industrial concepts.[102] In addition to utilizing many established features of robot controllers, such as position, velocity and force control of end effectors, they[clarification needed] also enable interconnection and the implementation of more advanced sensor fusion and control techniques, including adaptive control, fuzzy control and artificial neural network–based control.[102] When implemented in real time, such techniques can potentially enable more adaptive control systems when working in unfamiliar environments.[104] Generic reference architecture and associated interconnected, open-architecture robot and controller implementation has been used in a number of studies.[104][105]

    Sensing

    Main articles: Robotic sensing and Robotic sensors

    See also: Sensory-motor map

    A color sensor on a robot

    Sensors allow robots to receive information about the environment or internal components. This is essential for robots to perform their tasks and respond to changes with the appropriate response. Sensors are used for various forms of measurements, to provide real-time information, and to give the robots warnings; they can include cameras and microphones, as well as those that monitor network signals, power level, pressure, and temperature.

    Current robotic and prosthetic hands receive far less tactile information than the human hand. Recent research has developed a tactile sensor array that mimics the mechanical properties and touch receptors of human fingertips.[106][107] The sensor array is constructed as a rigid core surrounded by conductive fluid contained by an elastomeric skin. Electrodes are mounted on the surface of the rigid core and connected to an impedance-measuring device within the core. When the artificial skin touches an object, the fluid path around the electrodes is deformed, producing impedance changes that map the forces received from the object. An important function of artificial fingertips will likely be adjusting the grip on held objects. Scientists from several European countries and Israel developed a prosthetic hand in 2009 which functions like a real one—allowing patients to write, type on a keyboard, and perform other fine movements. The prosthesis has sensors which enable the patient to sense through its fingertips.[108]

    Other common forms of sensing in robotics use lidar, radar, and sonar.[74] Lidar measures the distance to a target by illuminating the target with laser light and measuring the reflected light with a sensor. Radar uses radio waves to determine the range, angle, or velocity of objects. Sonar uses sound propagation to navigate, communicate with or detect objects on or under the surface of the water. More abstractly, robot forms inspired by origami are designed to sense and analyze in extreme environments.[109]

    Cameras can capture visible light and other forms of electromagnetic radiation such as infrared. Multiple sensors and particular lenses may be used to achieve a certain field of view and depth perception. Computer vision has increasingly utilized machine learning.[110] Barcode scanning may be utilized, but is not necessarily universally effective.[111]

    Software

    Main article: Robot software

    TOPIO, a ping pong–playing robot

    A program is how a robot decides when or how to do something. Specific motions can be designed via computer animation software such as Autodesk Maya.[112] Programs can be run by remote control, AI, or a hybrid of the two.

    A robot with remote-control programming, possibly operated by haptic or teleoperated devices, has a preexisting set of commands that it will perform when it receives a signal from a control source—essentially a form of automation that humans have extensive control over.

    Meanwhile, AI-supported autonomous robots operate without a control source and use their programming to determine responses to various stimuli.[113] Highly predictable machines, such as industrial robots, typically do not require complex cognition.

    Hybrid robots may be assisted by an operator who selects certain modes or commands general actions, with AI determining the necessary motions.[114]

    Human–robot interaction

    Main article: Human–robot interaction

    For effective use in domestic environments, science-fiction authors and futurists often envision humans communicating with robots via speech, gestures, and facial expressions[115] rather than a command-line interface.[116][117] Sounds, facial expressions, and body language can be used to convey emotions.[118] Meanwhile, recognizing human speech in real time is a difficult task for a computer, mostly because of speech’s great variability.[119][120] AI-assisted machines can use voice to identify emotions.[121] Social robots will also likely need to be able to recognize gestures (and perhaps perform them) to assist verbal communication.[122][123]

    Developments in autonomy

    GPS, radar, and lidar are combined in a vehicle developed for 2007’s DARPA Urban Challenge.

    Robots that can operate autonomously in a dynamic environment require a combination of mapping and navigation hardware and software to traverse their environment and avoid colliding with other objects. Besides humanoids such as ASIMO and Meinü robots, this particularly applies to self-driving cars, which variously employ the Global Positioning System (GPS), radar, lidar, cameras, an inertial navigation system and/or swarms of autonomous robots.[124]

    By 2025, Figure AI developed a vision–language–action model that can control two humanoid robots at once.[125] By the next year, MIT developed a predictive technique allowing humanoid robots to think ahead instead of having to recalibrate in real-time as the environment changes; this increased their speed by two or three times while slightly decreasing their accuracy.[126]

    Applications

    Current and potential applications of robots include:

    Employment concerns

    A robot technician builds small all-terrain robots (courtesy: MobileRobots, Inc.).

    The incorporation of robots into industries has increased efficiency and productivity. It is typically seen as a long-term investment for benefactors and perhaps even an essential component of manufacturing. However, it has the potential of replacing most of the work performed by humans, with a 2017 study finding that automation alone puts 47% of US jobs at eventual risk.[143] Robotics is thus often used as an argument for basic income to replace lost wages.[144] Theoretical physicist Stephen Hawking observed in 2016:[145]

    The automation of factories has already decimated jobs in traditional manufacturing, and the rise of artificial intelligence is likely to extend this job destruction deep into the middle classes, with only the most caring, creative or supervisory roles remaining.

    According to the Robotic Industries Association US data, in 2016 the automotive industry was the main customer of industrial robots with 52% of total sales.[146] They can perform over half of the labor in the auto industry, including heavy duty such as car assembly.[147] By 2003, an IBM keyboard manufacturing factory in Texas was fully automated as a “lights out” factory.[148] In 2026, one expert observed that most humanoid logistics robots would likely be assigned to new, streamlined facilities, rather than those originally designed for humans.[149] As of 2022, China had the greatest number of industrial robots in operation with 1.5 million units and was increasing that figure by more than 20% annually.[150]

    Since 2018, autonomous restaurants have been run in Massachusetts (by Spyce Kitchen until 2022),[151] the Los Angeles area (by Miso Robotics), and the Chicago area (by Nala Robotics).[152] As of August 2026, California Pizza Kitchen plans to install 1,000 vending machines over three years.[153] Some companies sell their kitchen automation systems, including those related to commercial dishwashing.[137][154]

    In the first half of 2026, China delivered over 40,000 humanoid robots.[155] The Donald Trump administration subsequently banned the import of new advanced robot models from China, citing security concerns,[156] while US companies prepared to deliver some of the first robots intended for domestic use.[157][158]

    Safety and health

    Main article: Workplace robotics safety

    See also: Soft robotics

    Yamaha Motor‘s industrial cobot (collaborative robot)

    The spread of robotics presents both opportunities and challenges for occupational safety and health (OSH).[159][failed verification] Despite lost wages, the substitution of people working in unhealthy or dangerous environments is an OSH benefit. These include not only high-risk jobs in space,[citation needed] security,[160] and energy, but also dirty or unsafe work in logistics, maintenance, and inspection that require exposure to physical and/or psychosocial risks, including those stemming from repetitive or monotonous tasks better suited to machines. Robots are likely to gradually replace such jobs in other sectors like agriculture, cleaning, construction, firefighting, healthcare, and transportation.[additional citation(s) needed] In late 2026, InvestorPlace suggested that humanoid robots may be first deployed to work night shifts;[161] as of 2007 some 20% of the workforce of developed countries worked nights, but it likely contributes to the risk of cancer.[162][163][164]

    On the other hand, humans are better suited than machines for light-duty jobs involving various levels of creativity, decision-making, and flexibility. Humans and robots increasingly work in parallel within their areas of expertise. The need to work safely in a close space has resulted in cobots (collaborative robots).[165][166] Some European countries are including robotics in their national programs, promoting healthy cooperation between robots and operators to increase productivity.[167]

    Careers

    Robotics is an interdisciplinary field, primarily combining mechanical engineering and computer science but also drawing on electronic engineering and other subjects. Undergraduate degrees are usually obtained in one of these subjects prior to the pursuance of a graduate degree in robotics. Robotics engineers design and maintain robots, develop new applications, and conduct research.[168] As of 2011, the number of robotics-related jobs was steadily rising as factories increasingly utilized robots.[169] According to a September 2021 GlobalData report, the robotics industry was worth US$45 billion in 2020, and by 2030 it will have grown at a compound annual growth rate of 29% to US$568 bn, driving jobs in robotics and related industries.[170]

  • Computer vision

    Computer vision tasks include methods for acquiring, processing, analyzing, and understanding digital images, and extraction of high-dimensional data from the real world in order to produce numerical or symbolic information, e.g. in the form of decisions.[1][2][3][4] “Understanding” in this context signifies the transformation of visual images into descriptions of the world that make sense to thought processes and can elicit appropriate action. This image understanding can be seen as the disentangling of symbolic information from image data using models constructed with the aid of geometry, physics, statistics, and learning theory.

    The scientific discipline of computer vision is concerned with the theory behind artificial systems that extract information from images. Image data can take many forms, such as video sequences, views from multiple cameras, multi-dimensional data from a 3D scanner, 3D point clouds from LiDaR sensors, or medical scanning devices. The technological discipline of computer vision seeks to apply its theories and models to the construction of computer vision systems.

    Subdisciplines of computer vision include scene reconstruction, object detection, event detection, activity recognition, video tracking, object recognition, 3D pose estimation, learning, indexing, motion estimation, visual servoing, 3D scene modeling, and image restoration.

    Definition

    Computer vision is an interdisciplinary field that deals with how computers can be made to gain high-level understanding from digital images or videos. From the perspective of engineering, it seeks to automate tasks that the human visual system can do.[5][6][7] “Computer vision is concerned with the automatic extraction, analysis, and understanding of useful information from a single image or a sequence of images. It involves the development of a theoretical and algorithmic basis to achieve automatic visual understanding.”[8] As a scientific discipline, computer vision is concerned with the theory behind artificial systems that extract information from images. The image data can take many forms, such as video sequences, views from multiple cameras, or multi-dimensional data from a medical scanner.[9] As a technological discipline, computer vision seeks to apply its theories and models for the construction of computer vision systems. Machine vision refers to a systems engineering discipline, especially in the context of factory automation. In more recent times, the terms computer vision and machine vision have converged to a greater degree.[10]: 13 

    History

    In the late 1960s, computer vision began at universities that were pioneering artificial intelligence. It was meant to mimic the human visual system as a stepping stone to endowing robots with intelligent behavior.[11] In 1966, it was believed that this could be achieved through an undergraduate summer project,[12] by attaching a camera to a computer and having it “describe what it saw”.[13][14]

    What distinguished computer vision from the prevalent field of digital image processing at that time was a desire to extract three-dimensional structure from images with the goal of achieving full scene understanding. Studies in the 1970s formed the early foundations for many of the computer vision algorithms that exist today, including extraction of edges from images, labeling of lines, non-polyhedral and polyhedral modeling, representation of objects as interconnections of smaller structures, optical flow, and motion estimation.[11]

    The next decade saw studies based on more rigorous mathematical analysis and quantitative aspects of computer vision. These include the concept of scale-space, the inference of shape from various cues such as shading, texture and focus, and contour models known as snakes. Researchers also realized that many of these mathematical concepts could be treated within the same optimization framework as regularization and Markov random fields.[15] By the 1990s, some of the previous research topics became more active than others. Research in projective 3-D reconstructions led to better understanding of camera calibration. With the advent of optimization methods for camera calibration, it was realized that a lot of the ideas were already explored in bundle adjustment theory from the field of photogrammetry. This led to methods for sparse 3-D reconstructions of scenes from multiple images. Progress was made on the dense stereo correspondence problem and further multi-view stereo techniques. At the same time, variations of graph cut were used to solve image segmentation. This decade also marked the first time statistical learning techniques were used in practice to recognize faces in images (see Eigenface). Toward the end of the 1990s, a significant change came about with the increased interaction between the fields of computer graphics and computer vision. This included image-based rendering, image morphing, view interpolation, panoramic image stitching and early light-field rendering.[11]

    Recent work has seen the resurgence of feature-based methods used in conjunction with machine learning techniques and complex optimization frameworks.[16][17] The advancement of Deep Learning techniques has brought further life to the field of computer vision. The accuracy of deep learning algorithms on several benchmark computer vision data sets for tasks ranging from classification,[18] segmentation and optical flow has surpassed prior methods.[19][20]

    Related fields

    Object detection in a photograph

    Solid-state physics

    Solid-state physics is another field that is closely related to computer vision. Most computer vision systems rely on image sensors, which detect electromagnetic radiation, which is typically in the form of either visible, infrared or ultraviolet light. The sensors are designed using quantum physics. The process by which light interacts with surfaces is explained using physics. Physics explains the behavior of optics which are a core part of most imaging systems. Sophisticated image sensors even require quantum mechanics to provide a complete understanding of the image formation process.[11] Also, various measurement problems in physics can be addressed using computer vision, for example, motion in fluids.

    Neurobiology

    Neurobiology has greatly influenced the development of computer vision algorithms. Over the last century, there has been an extensive study of eyes, neurons, and brain structures devoted to the processing of visual stimuli in both humans and various animals. This has led to a coarse yet convoluted description of how natural vision systems operate in order to solve certain vision-related tasks. These results have led to a sub-field within computer vision where artificial systems are designed to mimic the processing and behavior of biological systems at different levels of complexity. Also, some of the learning-based methods developed within computer vision (e.g. neural net and deep learning based image and feature analysis and classification) have their background in neurobiology. The Neocognitron, a neural network developed in the 1970s by Kunihiko Fukushima, is an early example of computer vision taking direct inspiration from neurobiology, specifically the primary visual cortex, serving as a foundational architectural precursor to modern convolutional neural networks (CNNs).[21]

    Some strands of computer vision research are closely related to the study of biological vision—indeed, just as many strands of AI research are closely tied with research into human intelligence and the use of stored knowledge to interpret, integrate, and utilize visual information. The field of biological vision studies and models the physiological processes behind visual perception in humans and other animals. Computer vision, on the other hand, develops and describes the algorithms implemented in software and hardware behind artificial vision systems. An interdisciplinary exchange between biological and computer vision has proven fruitful for both fields.[22]

    • Simplified example of training a neural network in object detection: The network is trained by multiple images known to depict starfish and sea urchins, which are correlated with "nodes" representing visual features. Starfish match with a ringed texture and star outline, whereas most sea urchins match with a striped texture and oval shape.Simplified example of training a neural network in object detection: The network is trained by multiple images known to depict starfish and sea urchins, which are correlated with “nodes” representing visual features. Starfish match with a ringed texture and star outline, whereas most sea urchins match with a striped texture and oval shape.
    • The resulting model contains weighted associations between visual features and output categories. Because one training image depicts a ring-textured sea urchin, the model also develops a weak association between ringed texture and sea urchin.The resulting model contains weighted associations between visual features and output categories. Because one training image depicts a ring-textured sea urchin, the model also develops a weak association between ringed texture and sea urchin.
    • Subsequent run of the model on an input image (left): The network correctly detects the starfish. However, the weak association between ringed texture and sea urchin also gives a weak signal to the latter from one of two intermediate nodes. In addition, a shell not included in the training gives a weak signal for the oval shape, also resulting in a weak signal for the sea urchin output. These weak signals may result in a false positive result for sea urchin.Subsequent run of the model on an input image (left):[23] The network correctly detects the starfish. However, the weak association between ringed texture and sea urchin also gives a weak signal to the latter from one of two intermediate nodes. In addition, a shell not included in the training gives a weak signal for the oval shape, also resulting in a weak signal for the sea urchin output. These weak signals may result in a false positive result for sea urchin.

    Signal processing

    Yet another field related to computer vision is signal processing. Many methods for processing one-variable signals, typically temporal signals, can be extended in a natural way to the processing of two-variable signals or multi-variable signals in computer vision. However, because of the specific nature of images, there are many methods developed within computer vision that have no counterpart in the processing of one-variable signals. Together with the multi-dimensionality of the signal, this defines a subfield in signal processing as a part of computer vision.

    Robotic navigation

    Robot navigation sometimes deals with autonomous path planning or deliberation for robotic systems to navigate through an environment.[24] A detailed understanding of these environments is required to navigate through them. Information about the environment could be provided by a computer vision system, acting as a vision sensor and providing high-level information about the environment and the robot

    Visual computing

    This section is an excerpt from Visual computing.edit

    Visual computing is a generic term for all computer science disciplines dealing with images and 3D models, such as computer graphics, image processing, visualization, computer vision, virtual and augmented reality, video processing, and computational visualistics. Visual computing also includes aspects of pattern recognition, human computer interaction, machine learning and digital libraries. The core challenges are the acquisition, processing, analysis and rendering of visual information (mainly images and video). Application areas include industrial quality control, medical image processing and visualization, surveying, robotics, multimedia systems, virtual heritage, special effects in movies and television, and ludology. Visual computing also includes digital art and digital media studies.

    Other fields

    Besides the above-mentioned views on computer vision, many of the related research topics can also be studied from a purely mathematical point of view. For example, many methods in computer vision are based on statistics, optimization or geometry. Finally, a significant part of the field is devoted to the implementation aspect of computer vision; how existing methods can be realized in various combinations of software and hardware, or how these methods can be modified in order to gain processing speed without losing too much performance. Computer vision is also used in fashion eCommerce, inventory management, patent search, furniture, and the beauty industry.[25]

    Distinctions

    The fields most closely related to computer vision are image processing, image analysis and machine vision. There is a significant overlap in the range of techniques and applications that these cover. This implies that the basic techniques that are used and developed in these fields are similar, something which can be interpreted as there is only one field with different names. On the other hand, it appears to be necessary for research groups, scientific journals, conferences, and companies to present or market themselves as belonging specifically to one of these fields and, hence, various characterizations which distinguish each of the fields from the others have been presented. In image processing, the input and output are both images, whereas in computer vision, the input is an image or video, and the output could be an enhanced image, an analysis of the image’s content, or even a system’s behavior based on that analysis.

    Computer graphics produces image data from 3D models, and computer vision often produces 3D models from image data.[26] There is also a trend towards a combination of the two disciplines, e.g., as explored in augmented reality.

    The following characterizations appear relevant but should not be taken as universally accepted:

    • Image processing and image analysis tend to focus on 2D images, how to transform one image to another, e.g., by pixel-wise operations such as contrast enhancement, local operations such as edge extraction or noise removal, or geometrical transformations such as rotating the image. This characterization implies that image processing/analysis neither requires assumptions nor produces interpretations about the image content.
    • Computer vision includes 3D analysis from 2D images. This analyzes the 3D scene projected onto one or several images, e.g., how to reconstruct structure or other information about the 3D scene from one or several images. Computer vision often relies on more or less complex assumptions about the scene depicted in an image.
    • Machine vision is the process of applying a range of technologies and methods to provide imaging-based automatic inspection, process control, and robot guidance[27] in industrial applications.[22] Machine vision tends to focus on applications, mainly in manufacturing, e.g., vision-based robots and systems for vision-based inspection, measurement, or picking (such as bin picking[28]). This implies that image sensor technologies and control theory often are integrated with the processing of image data to control a robot and that real-time processing is emphasized by means of efficient implementations in hardware and software. It also implies that external conditions such as lighting can be and are often more controlled in machine vision than they are in general computer vision, which can enable the use of different algorithms.
    • There is also a field called imaging which primarily focuses on the process of producing images, but sometimes also deals with the processing and analysis of images. For example, medical imaging includes substantial work on the analysis of image data in medical applications. Progress in convolutional neural networks (CNNs) has improved the accurate detection of disease in medical images, particularly in cardiology, pathology, dermatology, and radiology.[29]
    • Finally, pattern recognition is a field that uses various methods to extract information from signals in general, mainly based on statistical approaches and artificial neural networks.[30] A significant part of this field is devoted to applying these methods to image data.

    Photogrammetry also overlaps with computer vision, e.g., stereophotogrammetry vs. computer stereo vision.

    Applications

    Applications range from tasks such as industrial machine vision systems which, say, inspect bottles speeding by on a production line, to research into artificial intelligence and computers or robots that can comprehend the world around them. The computer vision and machine vision fields have significant overlap. Computer vision covers the core technology of automated image analysis which is used in many fields. Machine vision usually refers to a process of combining automated image analysis with other methods and technologies to provide automated inspection and robot guidance in industrial applications. In many computer-vision applications, computers are pre-programmed to solve a particular task, but methods based on learning are now becoming increasingly common. Examples of applications of computer vision include systems for:

    Learning 3D shapes has been a challenging task in computer vision. Recent advances in deep learning have enabled researchers to build models that are able to generate and reconstruct 3D shapes from single or multi-view depth maps or silhouettes seamlessly and efficiently.[26]

    For 2024, the leading areas of computer vision were industry (market size US$5.22 billion),[35] medicine (market size US$2.6 billion),[36] military (market size US$996.2 million).[37]

    Medicine

    DARPA‘s Visual Media Reasoning concept video

    One of the most prominent application fields is medical computer vision, or medical image processing, characterized by the extraction of information from image data to diagnose a patient.[38] An example of this is the detection of tumours, arteriosclerosis or other malign changes, and a variety of dental pathologies; measurements of organ dimensions, blood flow, etc. are another example. It also supports medical research by providing new information: e.g., about the structure of the brain or the quality of medical treatments. Applications of computer vision in the medical area also include enhancement of images interpreted by humans—ultrasonic images or X-ray images, for example—to reduce the influence of noise.

    Machine vision

    A second application area in computer vision is in industry, sometimes called machine vision, where information is extracted for the purpose of supporting a production process. One example is quality control where details or final products are being automatically inspected in order to find defects. One of the most prevalent fields for such inspecti is the Wafer industry in which every single Wafer is being measured and inspected for inaccuracies or defects to prevent a computer chip from coming to market in an unusable manner. Another example is a measurement of the position and orientation of details to be picked up by a robot arm. Machine vision is also heavily used in the agricultural processes to remove undesirable foodstuff from bulk material, a process called optical sorting.[39]

    Military

    The obvious examples are the detection of enemy soldiers or vehicles and missile guidance. More advanced systems for missile guidance send the missile to an area rather than a specific target, and target selection is made when the missile reaches the area based on locally acquired image data. Modern military concepts, such as “battlefield awareness”, imply that various sensors, including image sensors, provide a rich set of information about a combat scene that can be used to support strategic decisions. In this case, automatic processing of the data is used to reduce complexity and to fuse information from multiple sensors to increase reliability.

    Autonomous vehicles

    Artist’s concept of Curiosity, an example of an uncrewed land-based vehicle. The stereo camera is mounted on top of the rover.

    One of the newer application areas is autonomous vehicles, which include submersibles, land-based vehicles (small robots with wheels, cars, or trucks), aerial vehicles, and unmanned aerial vehicles (UAV). The level of autonomy ranges from fully autonomous (unmanned) vehicles to vehicles where computer-vision-based systems support a driver or a pilot in various situations. Fully autonomous vehicles typically use computer vision for navigation, e.g., for knowing where they are or mapping their environment (SLAM), for detecting obstacles. It can also be used for detecting certain task-specific events, e.g., a UAV looking for forest fires. Examples of supporting systems are obstacle warning systems in cars, cameras and LiDAR sensors in vehicles, and systems for autonomous landing of aircraft. Several car manufacturers have demonstrated systems for autonomous driving of cars. There are ample examples of military autonomous vehicles ranging from advanced missiles to UAVs for recon missions or missile guidance. Space exploration is already being made with autonomous vehicles using computer vision, e.g., NASA‘s Curiosity and CNSA‘s Yutu-2 rover.

    Tactile feedback

    Rubber artificial skin layer with the flexible structure for the shape estimation of micro-undulation surfaces
    Above is a silicon mold with a camera inside containing many different point markers. When this sensor is pressed against the surface, the silicon deforms, and the position of the point markers shifts. A computer can then take this data and determine how exactly the mold is pressed against the surface. This can be used to calibrate robotic hands in order to make sure they can grasp objects effectively.

    Materials such as rubber and silicon are being used to create sensors that allow for applications such as detecting microundulations and calibrating robotic hands. Rubber can be used in order to create a mold that can be placed over a finger, inside of this mold would be multiple strain gauges. The finger mold and sensors could then be placed on top of a small sheet of rubber containing an array of rubber pins. A user can then wear the finger mold and trace a surface. A computer can then read the data from the strain gauges and measure if one or more of the pins are being pushed upward. If a pin is being pushed upward then the computer can recognize this as an imperfection in the surface. This sort of technology is useful in order to receive accurate data on imperfections on a very large surface.[40] Another variation of this finger mold sensor are sensors that contain a camera suspended in silicon. The silicon forms a dome around the outside of the camera and embedded in the silicon are point markers that are equally spaced. These cameras can then be placed on devices such as robotic hands in order to allow the computer to receive highly accurate tactile data.[41]

    Other applications

    Other application areas include:

    Typical tasks

    Each of the application areas described above employ a range of computer vision tasks; more or less well-defined measurement problems or processing problems, which can be solved using a variety of methods. Some examples of typical computer vision tasks are presented below.

    Computer vision tasks include methods for acquiring, processing, analyzing and understanding digital images, and extraction of high-dimensional data from the real world in order to produce numerical or symbolic information, e.g., in the forms of decisions.[1][2][3][4] Understanding in this context means the transformation of visual images into descriptions of the world that can interface with other thought processes and elicit appropriate action. This image understanding can be seen as the disentangling of symbolic information from image data using models constructed with the aid of geometry, physics, statistics, and learning theory.[47]

    Recognition

    The classical problem in computer vision, image processing, and machine vision is that of determining whether or not the image data contains some specific object, feature, or activity. Different varieties of recognition problem are described in the literature.[48]

    • Object recognition (also called object classification) – one or several pre-specified or learned objects or object classes can be recognized, usually together with their 2D positions in the image or 3D poses in the scene. Blippar, Google Goggles, and LikeThat provide stand-alone programs that illustrate this functionality.
    • Identification – an individual instance of an object is recognized. Examples include identification of a specific person’s face or fingerprint, identification of handwritten digits, or the identification of a specific vehicle.
    • Detection – the image data are scanned for specific objects along with their locations. Examples include the detection of an obstacle in the car’s field of view and possible abnormal cells or tissues in medical images or the detection of a vehicle in an automatic road toll system. Detection based on relatively simple and fast computations is sometimes used for finding smaller regions of interesting image data which can be further analyzed by more computationally demanding techniques to produce a correct interpretation.

    Currently, the best algorithms for such tasks are based on convolutional neural networks. An illustration of their capabilities is given by the ImageNet Large Scale Visual Recognition Challenge; this is a benchmark in object classification and detection, with millions of images and 1000 object classes used in the competition.[49] Performance of convolutional neural networks on the ImageNet tests is now close to that of humans.[49] The best algorithms still struggle with objects that are small or thin, such as a small ant on the stem of a flower or a person holding a quill in their hand. They also have trouble with images that have been distorted with filters (an increasingly common phenomenon with modern digital cameras). By contrast, those kinds of images rarely trouble humans. Humans, however, tend to have trouble with other issues. For example, they are not good at classifying objects into fine-grained classes, such as the particular breed of dog or species of bird, whereas convolutional neural networks handle this with ease.[citation needed]

    Several specialized tasks based on recognition exist, such as:

    • Content-based image retrieval – finding all images, which have a specific content, in a larger set of images. The content can be specified in different ways, for example in terms of similarity relative to a target image (give me all images similar to image X) by utilizing reverse image search techniques, or in terms of high-level search criteria given as text input (give me all images which contain many houses, are taken during winter and have no cars in them).
    Computer vision for people counter purposes in public places, malls, shopping centers
    • Pose estimation – estimating the position or orientation of a specific object relative to the camera. An example application for this technique would be assisting a robot arm in retrieving objects from a conveyor belt in an assembly line situation or picking parts from a bin.
    • Optical character recognition (OCR) – identifying characters in images of printed or handwritten text, usually with a view to encoding the text in a format more amenable to editing or indexing (e.g. ASCII). A related task is reading of 2D codes such as data matrix and QR codes.
    • Facial recognition – a technology that enables the matching of faces in digital images or video frames to a face database, which is now widely used for mobile phone facelock, smart door locking, etc.[50]
    • Emotion recognition – a subset of facial recognition, emotion recognition refers to the process of classifying human emotions. Psychologists caution, however, that internal emotions cannot be reliably detected from faces.[51]
    • Shape Recognition Technology (SRT) in people counter systems differentiating human beings (head and shoulder patterns) from objects.
    • Human activity recognition – deals with recognizing the activity from a series of video frames, such as, if the person is picking up an object or walking.

    Motion analysis

    Several tasks relate to motion estimation, where an image sequence is processed to produce an estimate of the velocity either at each points in the image or in the 3D scene or even of the camera that produces the images. Examples of such tasks are:

    • Egomotion – determining the 3D rigid motion (rotation and translation) of the camera from an image sequence produced by the camera.
    • Tracking – following the movements of a (usually) smaller set of interest points or objects (e.g., vehicles, objects, humans or other organisms[45]) in the image sequence. This has vast industry applications as most high-running machinery can be monitored in this way.
    • Optical flow – to determine, for each point in the image, how that point is moving relative to the image plane, i.e., its apparent motion. This motion is a result of both how the corresponding 3D point is moving in the scene and how the camera is moving relative to the scene.

    Scene reconstruction

    Given one or (typically) more images of a scene, or a video, scene reconstruction aims at computing a 3D model of the scene. In the simplest case, the model can be a set of 3D points. More sophisticated methods produce a complete 3D surface model. The advent of 3D imaging not requiring motion or scanning, and related processing algorithms is enabling rapid advances in this field. Grid-based 3D sensing can be used to acquire 3D images from multiple angles. Algorithms are now available to stitch multiple 3D images together into point clouds and 3D models.[26]

    Image restoration

    Image restoration comes into the picture when the original image is degraded or damaged due to some external factors like lens wrong positioning, transmission interference, low lighting or motion blurs, etc., which is referred to as noise. When the images are degraded or damaged, the information to be extracted from them also gets damaged. Therefore, we need to recover or restore the image as it was intended to be. The aim of image restoration is the removal of noise (sensor noise, motion blur, etc.) from images. The simplest possible approach for noise removal is various types of filters, such as low-pass filters or median filters. More sophisticated methods assume a model of how the local image structures look to distinguish them from noise. By first analyzing the image data in terms of the local image structures, such as lines or edges, and then controlling the filtering based on local information from the analysis step, a better level of noise removal is usually obtained compared to the simpler approaches.

    An example in this field is inpainting.

    System methods

    The organization of a computer vision system is highly application-dependent. Some systems are stand-alone applications that solve a specific measurement or detection problem, while others constitute a sub-system of a larger design which, for example, also contains sub-systems for control of mechanical actuators, planning, information databases, man-machine interfaces, etc. The specific implementation of a computer vision system also depends on whether its functionality is pre-specified or if some part of it can be learned or modified during operation. Many functions are unique to the application. There are, however, typical functions that are found in many computer vision systems.

    • Image acquisition – A digital image is produced by one or several image sensors, which, besides various types of light-sensitive cameras, include range sensors, tomography devices, radar, ultra-sonic cameras, etc. Depending on the type of sensor, the resulting image data is an ordinary 2D image, a 3D volume, or an image sequence. The pixel values typically correspond to light intensity in one or several spectral bands (gray images or colour images) but can also be related to various physical measures, such as depth, absorption or reflectance of sonic or electromagnetic waves, or magnetic resonance imaging.[39]
    • Pre-processing – Before a computer vision method can be applied to image data in order to extract some specific piece of information, it is usually necessary to process the data in order to ensure that it satisfies certain assumptions implied by the method. Examples are:
      • Re-sampling to ensure that the image coordinate system is correct.
      • Noise reduction to ensure that sensor noise does not introduce false information.
      • Contrast enhancement to ensure that relevant information can be detected.
      • Scale space representation to enhance image structures at locally appropriate scales.
    • Feature extraction – Image features at various levels of complexity are extracted from the image data.[39] Typical examples of such features are:

    More complex features may be related to texture, shape, or motion.

    • Detection/segmentation – At some point in the processing, a decision is made about which image points or regions of the image are relevant for further processing.[39] Examples are:
      • Selection of a specific set of interest points.
      • Segmentation of one or multiple image regions that contain a specific object of interest.
      • Segmentation of image into nested scene architecture comprising foreground, object groups, single objects or salient object[52] parts (also referred to as spatial-taxon scene hierarchy),[53] while the visual salience is often implemented as spatial and temporal attention.
      • Segmentation or co-segmentation of one or multiple videos into a series of per-frame foreground masks while maintaining its temporal semantic continuity.[54][55]
    • High-level processing – At this step, the input is typically a small set of data, for example, a set of points or an image region, which is assumed to contain a specific object.[39] The remaining processing deals with, for example:
      • Verification that the data satisfies model-based and application-specific assumptions.
      • Estimation of application-specific parameters, such as object pose or object size.
      • Image recognition – classifying a detected object into different categories.
      • Image registration – comparing and combining two different views of the same object.
    • Decision making – Making the final decision required for the application,[39] for example:
      • Pass/fail on automatic inspection applications.
      • Match/no-match in recognition applications.
      • Flag for further human review in medical, military, security and recognition applications.

    Image-understanding systems

    Image-understanding systems (IUS) include three levels of abstraction as follows: low level includes image primitives such as edges, texture elements, or regions; intermediate level includes boundaries, surfaces and volumes; and high level includes objects, scenes, or events. Many of these requirements are entirely topics for further research.

    The representational requirements in the designing of IUS for these levels are: representation of prototypical concepts, concept organization, spatial knowledge, temporal knowledge, scaling, and description by comparison and differentiation.

    While inference refers to the process of deriving new, not explicitly represented facts from currently known facts, control refers to the process that selects which of the many inference, search, and matching techniques should be applied at a particular stage of processing. Inference and control requirements for IUS are: search and hypothesis activation, matching and hypothesis testing, generation and use of expectations, change and focus of attention, certainty and strength of belief, inference and goal satisfaction.[56]

    Hardware

    A 2020 model iPad Pro with a LiDAR sensor

    There are many kinds of computer vision systems; however, all of them contain these basic elements: a power source, at least one image acquisition device (camera, ccd, etc.), a processor, and control and communication cables or some kind of wireless interconnection mechanism. In addition, a practical vision system contains software, as well as a display in order to monitor the system. Vision systems for inner spaces, as most industrial ones, contain an illumination system and may be placed in a controlled environment. Furthermore, a completed system includes many accessories, such as camera supports, cables, and connectors.

    Most computer vision systems use visible-light cameras passively viewing a scene at frame rates of at most 60 frames per second (usually far slower).

    A few computer vision systems use image-acquisition hardware with active illumination or something other than visible light or both, such as structured-light 3D scanners, thermographic cameras, hyperspectral imagers, radar imaging, lidar scanners, magnetic resonance images, side-scan sonar, synthetic aperture sonar, etc. Such hardware captures “images” that are then processed often using the same computer vision algorithms used to process visible-light images.

    While traditional broadcast and consumer video systems operate at a rate of 30 frames per second, advances in digital signal processing and consumer graphics hardware has made high-speed image acquisition, processing, and display possible for real-time systems on the order of hundreds to thousands of frames per second. For applications in robotics, fast, real-time video systems are critically important and often can simplify the processing needed for certain algorithms. When combined with a high-speed projector, fast image acquisition allows 3D measurement and feature tracking to be realized.[57]

    Egocentric vision systems are composed of a wearable camera that automatically take pictures from a first-person perspective.

    As of 2016, vision processing units are emerging as a new class of processors to complement CPUs and graphics processing units (GPUs) in this role.[58]

  • Natural Language Processing

    Natural language processing (NLP) is the processing of natural language information by a computer. NLP is a subfield of computer science and is closely associated with artificial intelligence. NLP is also related to information retrieval, knowledge representation, computational linguistics, and linguistics more broadly.[1]

    Major processing tasks in an NLP system include: speech recognition, text classification, natural language understanding, and natural language generation.

    History

    Further information: History of natural language processing

    Natural language processing has its roots in the 1950s.[2] Already in 1950, Alan Turing published an article titled “Computing Machinery and Intelligence,” which proposed what is now called the Turing test as a criterion of intelligence, though at the time that was not articulated as a problem separate from artificial intelligence. The proposed test includes a task that involves the automated interpretation and generation of natural language.

    Symbolic NLP (1950s – early 1990s)

    A document parsed into an abstract syntax tree

    The premise of symbolic NLP is often illustrated using John Searle’s Chinese room thought experiment: Given a collection of rules (e.g., a Chinese phrasebook, with questions and matching answers), the computer emulates natural language understanding (or other NLP tasks) by applying those rules to the data it confronts.

    • 1950s: The Georgetown experiment in 1954 involved fully automatic translation of more than sixty Russian sentences into English. The authors claimed that within three or five years, machine translation would be a solved problem.[3] However, real progress was much slower, and after the ALPAC report in 1966, which found that ten years of research had failed to fulfill the expectations, funding for machine translation was dramatically reduced. Little further research in machine translation was conducted in America (though some research continued elsewhere, such as Japan and Europe[4]) until the late 1980s when the first statistical machine translation systems were developed.
    • 1960s: Some notably successful natural language processing systems developed in the 1960s were SHRDLU, a natural language system working in restricted “blocks worlds” with restricted vocabularies, and ELIZA, a simulation of Rogerian psychotherapy, written by Joseph Weizenbaum between 1964 and 1966. Despite using minimal information about human thought or emotion, ELIZA was able to produce interactions that appeared human-like. When the “patient” exceeded the very small knowledge base, ELIZA might provide a generic response, for example, responding to “My head hurts” with “Why do you say your head hurts?”. Ross Quillian’s successful work on natural language was demonstrated with a vocabulary of only twenty words, because that was all that would fit in a computer memory at the time.[5]
    • 1970s: During the 1970s, many programmers began to write “conceptual ontologies“, which structured real-world information into computer-understandable data. Examples are MARGIE (Schank, 1975), SAM (Cullingford, 1978), PAM (Wilensky, 1978), TaleSpin (Meehan, 1976), QUALM (Lehnert, 1977), Politics (Carbonell, 1979), and Plot Units (Lehnert 1981). During this time, the first chatterbots were written (e.g., PARRY).
    • 1980s: The 1980s and early 1990s mark the heyday of symbolic methods in NLP. Focus areas of the time included research on rule-based parsing (e.g., the development of HPSG as a computational operationalization of generative grammar), morphology (e.g., two-level morphology[6]), semantics (e.g., Lesk algorithm), reference (e.g., within Centering Theory[7]) and other areas of natural language understanding (e.g., in the Rhetorical Structure Theory). Other lines of research were continued, e.g., the development of chatterbots with Racter and Jabberwacky. An important development (that eventually led to the statistical turn in the 1990s) was the rising importance of quantitative evaluation in this period.[8]

    Statistical NLP (1990s–present)

    Up until the 1980s, most natural language processing systems were based on complex sets of hand-written rules. Starting in the late 1980s, however, there was a revolution in natural language processing with the introduction of machine learning algorithms for language processing. This shift was influenced by increasing computational power (see Moore’s law) and a decline in the dominance of Chomskyan linguistic theories (e.g. transformational grammar), whose theoretical underpinnings discouraged the sort of corpus linguistics that underlies the machine-learning approach to language processing.[9]

    • 1990s: Many early successes of statistical NLP occurred in machine translation, particularly through work at IBM Research. An IBM team including Frederick Jelinek, Peter F. Brown, and Robert Mercer developed a probabilistic approach to translation, described in the 1990 paper A Statistical Approach to Machine Translation.[10] The work used parallel English and French text from the proceedings of the Parliament of Canada and was subsequently developed into the IBM alignment models.[11][12] However, such systems often depended on large task-specific corpora, making the availability of suitable training data an important limitation.
    • 2000s: With the growth of the web, increasing amounts of raw (unannotated) language data have become available since the mid-1990s. Research has thus increasingly focused on unsupervised and semi-supervised learning algorithms. Such algorithms can learn from data that has not been hand-annotated with the desired answers or using a combination of annotated and non-annotated data. Generally, this task is much more difficult than supervised learning, and typically produces less accurate results for a given amount of input data. However, large quantities of non-annotated data are available (including, among other things, the entire content of the World Wide Web), which can often make up for the worse efficiency if the algorithm used has a low enough time complexity to be practical.
    • 2003: word n-gram model, at the time the best statistical algorithm, is outperformed by a multi-layer perceptron (with a single hidden layer and context length of several words, trained on up to 14 million words, by Bengio et al.)[13]
    • 2010: Tomáš Mikolov (then a PhD student at Brno University of Technology) with co-authors applied a simple recurrent neural network with a single hidden layer to language modeling,[14] and in the following years he went on to develop Word2vec. In the 2010s, representation learning and deep neural network-style (featuring many hidden layers) machine learning methods became widespread in natural language processing. This shift gained momentum due to results showing that such techniques[15][16] can achieve state-of-the-art results in many natural language tasks, e.g., in language modeling[17] and parsing.[18][19] This is increasingly important in medicine and healthcare, where NLP helps analyze notes and text in electronic health records that would otherwise be inaccessible for study when seeking to improve care[20] or protect patient privacy.[21]

    Approaches: Symbolic, statistical, neural networks

    Symbolic approach, i.e., the hand-coding of a set of rules for manipulating symbols, coupled with a dictionary lookup, was historically the first approach used both by AI in general and by NLP in particular:[22][23] such as by writing grammars or devising heuristic rules for stemming.

    Machine learning approaches, which include both statistical and neural networks, on the other hand, have many advantages over the symbolic approach:

    • both statistical and neural network methods tend to focus more on the most common cases extracted from a corpus of texts, whereas the rule-based approach needs to provide rules for both rare and common cases equally.
    • language models, produced by either statistical or neural network methods, are more robust to both unfamiliar (e.g. containing words or structures that have not been seen before) and erroneous input (e.g. with misspelled words or words accidentally omitted) in comparison to the rule-based systems, which are also more costly to produce.
    • the larger such a (probabilistic) language model is, the more accurate it becomes, in contrast to rule-based systems that can gain accuracy only by increasing the amount and complexity of the rules leading to intractability problems.

    Rule-based systems are commonly used:

    • when the amount of training data is insufficient to successfully apply machine learning methods, e.g., for the machine translation of low-resource languages such as provided by the Apertium system,
    • for preprocessing in NLP pipelines, e.g., tokenization, or
    • for post-processing and transforming the output of NLP pipelines, e.g., for knowledge extraction from syntactic parses.

    Statistical approach

    In the late 1980s and mid-1990s, the statistical approach ended a period of AI winter, which was caused by the inefficiencies of the rule-based approaches.[24][25]

    The earliest decision trees, producing systems of hard if–then rules, were still very similar to the old rule-based approaches. Only the introduction of hidden Markov models, applied to part-of-speech tagging, announced the end of the old rule-based approach.

    Neural networks

    Further information: Artificial neural network

    A major drawback of statistical methods is that they require elaborate feature engineering. Since 2015,[26] neural network–based methods have increasingly replaced traditional statistical approaches, using semantic networks[27] and word embeddings to capture semantic properties of words.

    Intermediate tasks (e.g., part-of-speech tagging and dependency parsing) are not needed anymore.

    Neural machine translation, based on the then-newly invented sequence-to-sequence transformations, made obsolete the intermediate steps, such as word alignment, previously necessary for statistical machine translation.

    Common NLP tasks

    The following is a list of some of the most commonly researched tasks in natural language processing. Some of these tasks have direct real-world applications, while others more commonly serve as subtasks that are used to aid in solving larger tasks.

    Though natural language processing tasks are closely intertwined, they can be subdivided into categories for convenience. A coarse division is given below.

    Text and speech processing

    Word cloud of stop words in Hebrew

    Optical character recognition (OCR)Given an image representing printed text, determine the corresponding text.Speech recognitionGiven a sound clip of a person or people speaking, determine the textual representation of the speech. This is the opposite of text to speech and is one of the extremely difficult problems colloquially termed “AI-complete” (see above). In natural speech there are hardly any pauses between successive words, and thus speech segmentation is a necessary subtask of speech recognition (see below). In most spoken languages, the sounds representing successive letters blend into each other in a process termed coarticulation, so the conversion of the analog signal to discrete characters can be a very difficult process. Also, given that words in the same language are spoken by people with different accents, the speech recognition software must be able to recognize the wide variety of input as being identical to each other in terms of its textual equivalent.Speech segmentationGiven a sound clip of a person or people speaking, separate it into words. A subtask of speech recognition and typically grouped with it.Text-to-speechGiven a text, transform those units and produce a spoken representation. Text-to-speech can be used to aid the visually impaired.[28]Word segmentation (Tokenization)

    Tokenization is a text-processing technique that divides text into individual words or word fragments. This technique results in two key components: a word index and tokenized text. The word index is a list that maps unique words to specific numerical identifiers, and the tokenized text replaces each word with its corresponding numerical token. These numerical tokens are then used in various deep learning methods.[29]

    For a language like English, this is fairly trivial, since words are usually separated by spaces. However, some written languages like Chinese, Japanese and Thai do not mark word boundaries in such a fashion, and in those languages text segmentation is a significant task requiring knowledge of the vocabulary and morphology of words in the language. Sometimes this process is also used in cases like bag of words (BOW) creation in data mining.[30][31]

    Morphological analysis

    Lemmatization of Basque words

    LemmatizationThe task of removing inflectional endings only and returning the base dictionary form of a word which is also known as a lemma. Lemmatization is another technique for reducing words to their normalized form. But in this case, the transformation actually uses a dictionary to map words to their actual form.[32]Morphological segmentationSeparate words into individual morphemes and identify the class of the morphemes. The difficulty of this task depends greatly on the complexity of the morphology (i.e., the structure of words) of the language being considered. English has fairly simple morphology, especially inflectional morphology, and thus it is often possible to ignore this task entirely and simply model all possible forms of a word (e.g., “open, opens, opened, opening”) as separate words. In languages such as Turkish or Meitei, a highly agglutinated Indian language, however, such an approach is not possible, as each dictionary entry has thousands of possible word forms.[33]Part-of-speech taggingGiven a sentence, determine the part of speech (POS) for each word. Many words, especially common ones, can serve as multiple parts of speech. For example, “book” can be a noun (“the book on the table”) or verb (“to book a flight”); “set” can be a noun, verb or adjective; and “out” can be any of at least five different parts of speech.StemmingThe process of reducing inflected (or sometimes derived) words to a base form (e.g., “close” will be the root for “closed”, “closing”, “close”, “closer” etc.). Stemming yields similar results as lemmatization, but does so on grounds of rules, not a dictionary.

    Syntactic analysis

    Part of a series on
    Formal languages
    Key conceptsFormal systemAlphabetSyntaxFormal semanticsSemantics (programming languages)Formal grammarFormation ruleWell-formed formulaAutomata theoryRegular expressionProductionGround expressionAtomic formula
    Applications
    vte

    Grammar induction[34]Generate a formal grammar that describes a language’s syntax.Sentence breaking (also known as “sentence boundary disambiguation“)Given a chunk of text, find the sentence boundaries. Sentence boundaries are often marked by periods or other punctuation marks, but these same characters can serve other purposes (e.g., marking abbreviations).ParsingDetermine the parse tree (grammatical analysis) of a given sentence. The grammar for natural languages is ambiguous and typical sentences have multiple possible analyses: perhaps surprisingly, for a typical sentence there may be thousands of potential parses (most of which will seem completely nonsensical to a human). There are two primary types of parsing: dependency parsing and constituency parsing. Dependency parsing focuses on the relationships between words in a sentence (marking things like primary objects and predicates), whereas constituency parsing focuses on building out the parse tree using a probabilistic context-free grammar (PCFG) (see also stochastic grammar).

    Lexical semantics (of individual words in context)

    An entity linking pipeline

    Lexical semanticsWhat is the computational meaning of individual words in context?Distributional semanticsHow can we learn semantic representations from data?Named entity recognition (NER)Given a stream of text, determine which items in the text map to proper names, such as people or places, and what the type of each such name is (e.g. person, location, organization). Although capitalization can aid in recognizing named entities in languages such as English, this information cannot aid in determining the type of named entity, and in any case, is often inaccurate or insufficient. For example, the first letter of a sentence is also capitalized, and named entities often span several words, only some of which are capitalized. Furthermore, many other languages in non-Western scripts (e.g. Chinese or Arabic) do not have any capitalization at all, and even languages with capitalization may not consistently use it to distinguish names. For example, German capitalizes all nouns, regardless of whether they are names, and French and Spanish do not capitalize names that serve as adjectives. This task is also referred to as token classification.[35]Sentiment analysis (see also Multimodal sentiment analysis)Sentiment analysis involves identifying and classifying the emotional tone expressed in text. This technique involves analyzing text to determine whether the expressed sentiment is positive, negative, or neutral. Models for sentiment classification typically utilize inputs such as word n-grams, Term Frequency-Inverse Document Frequency (TF-IDF) features, hand-generated features, or employ deep learning models designed to recognize both long-term and short-term dependencies in text sequences. The applications of sentiment analysis are diverse, extending to tasks such as categorizing customer reviews on various online platforms.[29]Terminology extractionThe goal of terminology extraction is to automatically extract relevant terms from a given corpus.Word-sense disambiguation (WSD)Many words have more than one meaning; we have to select the meaning which makes the most sense in context. For this problem, we are typically given a list of words and associated word senses, e.g. from a dictionary or an online resource such as WordNet.Entity linkingMany words—typically proper names—refer to named entities; here we have to select the entity (a famous individual, a location, a company, etc.) which is referred to in context.

    Relational semantics (semantics of individual sentences)

    Relationship extractionGiven a chunk of text, identify the relationships among named entities (e.g. who is married to whom).Semantic parsingGiven a piece of text (typically a sentence), produce a formal representation of its semantics, either as a graph (e.g., in AMR parsing) or in accordance with a logical formalism (e.g., in DRT parsing). This challenge typically includes aspects of several more elementary NLP tasks from semantics (e.g., semantic role labelling, word-sense disambiguation) and can be extended to include full-fledged discourse analysis (e.g., discourse analysis, coreference; see Natural language understanding below).Semantic role labelling (see also implicit semantic role labelling below)Given a single sentence, identify and disambiguate semantic predicates (e.g., verbal frames), then identify and classify the frame elements (semantic roles).

    Discourse (semantics beyond individual sentences)

    Coreference resolutionGiven a sentence or larger chunk of text, determine which words (“mentions”) refer to the same objects (“entities”). Anaphora resolution is a specific example of this task, and is specifically concerned with matching up pronouns with the nouns or names to which they refer. The more general task of coreference resolution also includes identifying so-called “bridging relationships” involving referring expressions. For example, in a sentence such as “He entered John’s house through the front door”, “the front door” is a referring expression and the bridging relationship to be identified is the fact that the door being referred to is the front door of John’s house (rather than of some other structure that might also be referred to).Discourse analysisThis rubric includes several related tasks. One task is discourse parsing, i.e., identifying the discourse structure of a connected text, i.e. the nature of the discourse relationships between sentences (e.g. elaboration, explanation, contrast). Another possible task is recognizing and classifying the speech acts in a chunk of text (e.g. yes–no question, content question, statement, assertion, etc.).Implicit semantic role labellingGiven a single sentence, identify and disambiguate semantic predicates (e.g., verbal frames) and their explicit semantic roles in the current sentence (see Semantic role labelling above). Then, identify semantic roles that are not explicitly realized in the current sentence, classify them into arguments that are explicitly realized elsewhere in the text and those that are not specified, and resolve the former against the local text. A closely related task is zero anaphora resolution, i.e., the extension of coreference resolution to pro-drop languages.Recognizing textual entailmentGiven two text fragments, determine if one being true entails the other, entails the other’s negation, or allows the other to be either true or false.[36]Topic segmentation and recognitionGiven a chunk of text, separate it into segments each of which is devoted to a topic, and identify the topic of the segment.Argument miningThe goal of argument mining is the automatic extraction and identification of argumentative structures from natural language text with the aid of computer programs.[37] Such argumentative structures include the premise, conclusions, the argument scheme and the relationship between the main and subsidiary argument, or the main and counter-argument within discourse.[38][39]

    Higher-level NLP applications

    Machine translation in Firefox

    Automatic summarization (text summarization)Produce a readable summary of a chunk of text. Often used to provide summaries of the text of a known type, such as research papers, articles in the financial section of a newspaper.Grammatical error correctionGrammatical error detection and correction involves a great bandwidth of problems on all levels of linguistic analysis (phonology/orthography, morphology, syntax, semantics, pragmatics). Grammatical error correction is impactful since it affects hundreds of millions of people who use or acquire English as a second language. It has thus been subject to a number of shared tasks since 2011.[40][41][42] As far as orthography, morphology, syntax and certain aspects of semantics are concerned, and due to the development of powerful neural language models such as GPT-2, this can now (2019) be considered a largely solved problem and is being marketed in various commercial applications.Logic translationTranslate a text from a natural language into formal logic.Machine translation (MT)Automatically translate text from one human language to another. This is one of the most difficult problems, and is a member of a class of problems colloquially termed “AI-complete“, i.e. requiring all of the different types of knowledge that humans possess (grammar, semantics, facts about the real world, etc.) to solve properly.Natural language understanding (NLU)Convert chunks of text into more formal representations such as first-order logic structures that are easier for computer programs to manipulate. Natural language understanding involves the identification of the intended semantic from the multiple possible semantics that can be derived from a natural language expression which usually takes the form of organized notations of natural language concepts. Introduction and creation of language metamodels and ontologies are efficient but empirical solutions. An explicit formalization of natural language semantics without confusion with implicit assumptions such as closed-world assumption (CWA) vs. open-world assumption, or subjective Yes/No vs. objective True/False is expected for the construction of a basis of semantics formalization.[43]Natural language generation (NLG):Convert information from computer databases or semantic intents into readable human language.Book generationNot an NLP task proper but an extension of natural language generation and other NLP tasks is the creation of full-fledged books. The first machine-generated book was created by a rule-based system in 1984 (Racter, The policeman’s beard is half-constructed).[44] The first published work by a neural network was published in 2018, 1 the Road, marketed as a novel, contains sixty million words. Both these systems are basically elaborate but nonsensical (semantics-free) language models. The first machine-generated science book was published in 2019 (Beta Writer, Lithium-Ion Batteries, Springer, Cham).[45] Unlike Racter and 1 the Road, this is grounded on factual knowledge and based on text summarization.Document AIA Document AI platform sits on top of the NLP technology enabling users with no prior experience of artificial intelligence, machine learning or NLP to quickly train a computer to extract the specific data they need from different document types. NLP-powered Document AI enables non-technical teams to quickly access information hidden in documents, for example, lawyers, business analysts and accountants.[46]Dialogue managementComputer systems intended to converse with a human.Question answeringGiven a human-language question, determine its answer. Typical questions have a specific right answer (such as “What is the capital of Canada?”), but sometimes open-ended questions are also considered (such as “What is the meaning of life?”).Text-to-image generationGiven a description of an image, generate an image that matches the description.[47]Text-to-scene generationGiven a description of a scene, generate a 3D model of the scene.[48][49]Text-to-videoGiven a description of a video, generate a video that matches the description.[50][51]

    General tendencies and (possible) future directions

    Based on long-standing trends in the field, it is possible to extrapolate future directions of NLP. As of 2020, three trends among the topics of the long-standing series of CoNLL Shared Tasks can be observed:[52]

    • Interest in increasingly abstract, “cognitive” aspects of natural language (1999–2001: shallow parsing, 2002–03: named entity recognition, 2006–09/2017–18: dependency syntax, 2004–05/2008–09 semantic role labelling, 2011–12 coreference, 2015–16: discourse parsing, 2019: semantic parsing).
    • Increasing interest in multilinguality, and, potentially, multimodality (English since 1999; Spanish, Dutch since 2002; German since 2003; Bulgarian, Danish, Japanese, Portuguese, Slovenian, Swedish, Turkish since 2006; Basque, Catalan, Chinese, Greek, Hungarian, Italian, Turkish since 2007; Czech since 2009; Arabic since 2012; 2017: 40+ languages; 2018: 60+/100+ languages)
    • Elimination of symbolic representations (rule-based over supervised towards weakly supervised methods, representation learning and end-to-end systems)

    Cognition

    Most higher-level NLP applications involve aspects that emulate intelligent behavior and apparent comprehension of natural language. More broadly speaking, the technical operationalization of increasingly advanced aspects of cognitive behavior represents one of the developmental trajectories of NLP(see trends among CoNLL shared tasks above).

    Cognition refers to “the mental action or process of acquiring knowledge and understanding through thought, experience, and the senses.”[53] Cognitive science is the interdisciplinary, scientific study of the mind and its processes.[54] Cognitive linguistics is an interdisciplinary branch of linguistics, combining knowledge and research from both psychology and linguistics.[55] Especially during the age of symbolic NLP, the area of computational linguistics maintained strong ties with cognitive studies.

    As an example, George Lakoff offers a methodology to build natural language processing (NLP) algorithms through the perspective of cognitive science, along with the findings of cognitive linguistics,[56] with two defining aspects:

    1. Apply the theory of conceptual metaphor, explained by Lakoff as “the understanding of one idea, in terms of another” which provides an idea of the intent of the author.[57] For example, consider the English word big. When used in a comparison (“That is a big tree”), the author intends to imply that the tree is physically large relative to other trees or the author’s experience. When used metaphorically (“Tomorrow is a big day”), the author intends to imply importance. The intent behind other usages, like in “She is a big person”, will remain somewhat ambiguous to a person and a cognitive NLP algorithm alike without additional information.
    2. Assign relative measures of meaning to a word, phrase, sentence or piece of text based on the information presented before and after the piece of text being analyzed, e.g., by means of a probabilistic context-free grammar (PCFG). The mathematical equation for such algorithms is presented in US Patent 9269353:[58]RMM(tokenN)=PMM(tokenN)×12d(∑i=−dd((PMM(tokenN)×PF(tokenN−i,tokenN,tokenN+i))i){\displaystyle {RMM(token_{N})}={PMM(token_{N})}\times {\frac {1}{2d}}\left(\sum _{i=-d}^{d}{((PMM(token_{N})}\times {PF(token_{N-i},token_{N},token_{N+i}))_{i}}\right)}where
      • RMM is the relative measure of meaning
      • token is any block of text, sentence, phrase or word
      • N is the number of tokens being analyzed
      • PMM is the probable measure of meaning based on a corpora
      • d is the non-zero location of the token along the sequence of N tokens
      • PF is the probability function specific to a language

    Ties with cognitive linguistics are part of the historical heritage of NLP, but they have been less frequently addressed since the statistical turn during the 1990s. Nevertheless, approaches to develop cognitive models towards technically operationalizable frameworks have been pursued in the context of various frameworks, e.g., of cognitive grammar,[59] functional grammar,[60] construction grammar,[61] computational psycholinguistics and cognitive neuroscience (e.g., ACT-R), however, with limited uptake in mainstream NLP (as measured by presence on major conferences[62] of the ACL). More recently, ideas of cognitive NLP have been revived as an approach to achieve explainability, e.g., under the notion of “cognitive AI”.[63] Likewise, ideas of cognitive NLP are inherent to neural models multimodal NLP (although rarely made explicit)[64] and developments in artificial intelligence, specifically tools and technologies using large language model approaches[65] and new directions in artificial general intelligence based on the free energy principle[66] by British neuroscientist and theoretician at University College London Karl J. Friston.