Machine unlearning is a branch of machine learning focused on removing specific undesired element, such as private data, wrong or manipulated training data, outdated information, copyrighted material, harmful content, dangerous abilities, or misinformation, without needing to rebuild models from the ground up. Large language models, like the ones powering ChatGPT, may be asked not just to remove specific elements but also to unlearn a "concept," "fact," or "knowledge," which aren't easily linked to specific examples. New terms such as "model editing," "concept editing," and "knowledge unlearning" have emerged to describe this process. == History == Early research efforts were largely motivated by Article 17 of the GDPR, the European Union's privacy regulation commonly known as the "right to be forgotten" (RTBF), introduced in 2014. The GDPR did not anticipate that the development of large language models would make data erasure a complex task. This issue has since led to research on "machine unlearning," with a growing focus on removing copyrighted material, harmful content, dangerous capabilities, and misinformation. Just as early experiences in humans shape later ones, some concepts are more fundamental and harder to unlearn. A piece of knowledge may be so deeply embedded in the model's knowledge graph that unlearning it could cause internal contradictions, requiring adjustments to other parts of the graph to resolve them. Researchers have now also started studying unlearning in the context of removing incorrect or adversarially manipulated training data such as systematically biased labels or poisoning attacks. == Motivations == At present, machine unlearning is motivated by a growing range of concerns that extend well beyond the field's original focus on data privacy. A widely used taxonomy in the literature distinguishes two high-level categories of motivation. Access revocation covers cases where a data subject or rights holder requests the removal of data they own or control. This is most commonly associated with RTBF established by the European Union's General Data Protection Regulation (GDPR) and analogous legislation such as the California Consumer Privacy Act (CCPA). These regulations grant individuals the legal right to request erasure of their personal data from any system that has processed it, including models that were trained on it. Access revocation also encompasses the removal of copyrighted or pay-walled content that was incorporated into training corpora without the necessary licenses, a concern that has become prominent with the widespread use of largely web-scraped pre-training datasets. Model correction covers cases where the model exhibits undesirable behavior arising from the training data, regardless of any individual's request. This includes: Removal of toxic, biased, or unsafe outputs introduced by harmful content in the training set Correction of stale or factually incorrect associations, such as outdated knowledge encoded in a deployed model Removal of dangerous capabilities, such as detailed knowledge of the synthesis of chemical or biological agents Correction of the influence of data poisoning or adversarial attacks that have corrupted model behavior This second category has been formalized as corrective machine unlearning, which frames unlearning as a post-training mechanism for repairing the effects of bad or harmful training data. It is closely related to the AI safety literature, where data filtering alone has been found insufficient to prevent hazardous knowledge from being encoded in model weights, motivating unlearning as a complementary risk mitigation strategy. A further distinction has been drawn in the literature between removal {eliminating the influence of specific training data on model parameters) and suppression (preventing the model from generating specific outputs regardless of how that knowledge is encoded). These two goals are not equivalent: removing training data does not guarantee meaningful output suppression, and suppressing outputs does not constitute removal of the underlying training data's influence. == SISA Training == SISA is a training strategy consisting of four mechanisms designed to make machine unlearning more efficient by structuring how models are trained and updated. Its goal is to allow a system to remove the influence of specific data points without retraining an entire model from scratch. By reorganizing training data and workflows, SISA reduces the computational burden of unlearning requests. Sharding divides the training dataset into multiple disjoint subsets, or shards. Each shard is used to train a separate model instance. This ensures that a single data point affects only one shard, so unlearning it requires updating only the corresponding shard rather than the full model. Isolation refers to training each shard independently, with nothing shared across shards during the training process. This separation prevents cross-contamination between shards, ensuring that forgetting data in one shard does not require adjustments to any others. Slicing breaks the data within each shard into sequential slices and stores model states after each slice is trained on. When an unlearning request targets a piece of data, the system can roll back to the checkpoint before the point was seen and retrain only from that slice forward. This reduces retraining time even within a shard. Aggregation occurs at inference, when the model is queried. It combines the outputs of each shard to determine the output of the overall model. This is often through majority voting or averaging. This allows SISA-trained systems to behave like a single model despite being composed of multiple shard-level models. Together, these mechanisms enable machine learning systems to forget specific data points with far lower computational cost than full retraining. The trade-off is that sharding and slicing can lead to reduced model accuracy, worse generalization, and increased storage requirements for the intermediate checkpoints. This can be tolerable based on the needs of the individual or organization to comply with "right to be forgotten" or efficiently recover from backdoor attacks. == Algorithms == Machine unlearning algorithms are broadly categorized into exact and approximate methods, reflecting a fundamental trade-off between formal guarantees and computational tractability. === Exact Unlearning === Exact unlearning methods produce a model that is statistically indistinguishable from one retrained from scratch on the dataset with the forget data removed. The canonical framework for exact unlearning is SISA Training (Sharded, Isolated, Sliced, and Aggregated), introduced by Bourtoule et al. (2021). SISA partitions the training dataset into disjoint shards and trains a separate sub-model on each. At inference time, predictions are aggregated across sub-models. When an unlearning request is received, only the sub-model corresponding to the shard containing the target data requires retraining, reducing computational overhead proportionally to the number of shards. Exact methods provide the strongest guarantees but become prohibitively expensive for large pre-trained neural networks and are generally limited to settings where training can be structured in advance. === Approximate Unlearning === Approximate unlearning methods seek to produce a model whose behavior is sufficiently close to an exactly unlearned model without the cost of full retraining. These methods dominate practical applications. Common approaches include: Gradient Ascent: The model is fine-tuned by maximizing the loss on the forget set, directly degrading its performance on targeted data. This is the most direct approach but risks destabilizing performance on retained data. Random Labelling: The model is fine-tuned on the forget set using randomly shuffled labels, confusing its associations with the targeted data while producing a less aggressive weight shift than pure gradient ascent. Gradient Difference: Combines gradient ascent on the forget set with simultaneous gradient descent on the retain set, using the retain objective as a regularizer to preserve general model utility. KL Divergence Regularization: Minimizes the KL divergence between the outputs of the unlearned model and the original model on the retain set, anchoring behavior on data the model should remember. Weight Pruning and Fine-tuning: Parameters with the smallest L1-norm are pruned — targeting weights most weakly associated with general knowledge and potentially most associated with the forget set — followed by fine-tuning on the retain set to restore utility. Layer Reset and Fine-tuning: The first or last k layers are re-initialized to random weights and the model is subsequently fine-tuned on the retain set. This is a coarse but computationally simple approach. Selective Synaptic Dampening: Uses influence functions to estimate the effect of individual trainin
SWILE
SWILE (formerly: Lunchr) is a French app-based company that focuses on improving the employee experience. Among others, the platform offers meal vouchers, gift vouchers, mobility vouchers, and business travel solutions. In March 2020, it was renamed SWILE and entered the lunch break and meal voucher market. == History == The company was founded as Lunchr by Loïc Soubeyrand in 2016. Originally, Lunchr was an app for pre-ordering lunch on the spot or to go. In January 2017, the company raised €2.5 million in seed funding from Daphni. In 2018, the company raised €11 million (series A) from Idinvest, followed by another €30 million in February 2019 (series B), notably from Index Ventures and Kima Ventures. In January 2020, Lunchr became one of the first startups to join the French Tech 120. A few months later, in March, Lunchr diversified its services, adding team life management tools and changing its brand name to Swile. In June 2020, the company raised €70 million more in a new round of financing (Series C) from the same investors and the BPI. In November 2020, Swile acquired Briq, a startup specializing in employee engagement. In January 2021, Swile won a tender with Carrefour and distributed 62,000 Swile cards to its employees. In early October 2021, a new $200 million (€175 million) fundraising round, in which Japanese Softbank joined other investors, allowed Swile to capitalize on $1 billion. President Emmanuel Macron cited the company as "a further proof that FrenchTech is at the forefront internationally." In May 2022, the company acquired the travel management start-up Okarito for €6 million. == Overview == Swile operates in two countries (France and Brazil) and has a total of 1000 employees, 5.5 million users and 85,000 corporate customers, including Carrefour, Le Monde, JCDECAUX, PSG, Airbnb, Spotify, Red Bull, and TikTok in the private sector, as well as numerous local authorities and ministerial references in the public sector.
LiveChat
LiveChat is an AI customer service software with chatbot, online chat, help desk software, and web analytics capabilities. LiveChat is used by over 76,000 companies. It was first launched in 2002 and is offered via a SaaS (software as a service) business model by Text. Organizations use LiveChat as a single point of contact to manage customer service and online sales activities with a single program. == Product == LiveChat is proprietary software. LiveChat's website chat widget can be embedded on customers' websites as a small chat box, often displayed in the bottom right corner of the web browser. It can be used to conduct chats, share files and save transcripts. The agent application is used by company employees to respond to questions asked by the customers. This is available through both web-based application, desktop applications, and mobile apps. Web chat sessions can be initiated by the visiting customer, or by the agent, either manually or automatically by the LiveChat system when the visitor meets the predefined criteria (i.e. searched keyword, time on website, encountered error, etc.). LiveChat's system attempts to identify the best prospects visiting a website based on data gathered from past purchasing decisions. Other features include real-time website traffic monitoring, built-in ticketing system and agents' efficiency analytics. LiveChat is available in 48 languages. == Research and reception == Reviewing LiveChat's usefulness for online learning in 2020, psychologist Jaclyn Broadbent said "LiveChat occurs as a real-time conversation, it can be time-consuming for staff and disruptive to other tasks." However, using it has resulted in reduced communication traffic from other channels, such as the discussion boards or email. As a teacher, the best time to be available on LiveChat is when you are doing other administrative jobs." Since 2014 LiveChat has been publishing Customer Service Report - an annual study of customer satisfaction and analysis of online business communication trends. It includes research of thousands of companies and millions of customer service email and live support interactions.
EasyChair
EasyChair is a web-based conference management software system. It has been used since 2002 in the scientific community for tasks such as organising research paper submission and review. In 2012, EasyChair added an open access online publication service for conference proceedings. == Description == EasyChair is a paid web-based conference management software system used, among other tasks, to organize paper submission and review, similar to other event management system software such as OpenConf. EasyChair used to be run by the Department of Computer Science at the University of Manchester but now it is a commercial service, owned by EasyChair Ltd. in Stockport (established 2016). EasyChair used to be free, for standard service, but as of 2022, only minimal services are free. The EasyChair website also provides an open access online publication service for conference proceedings. When launched in 2012, the service was for computer science only, but in 2016 it was expanded to all sciences. == History == The EasyChair software has been in continuous development since 2002. As of 2015, the code base consists of nearly 300,000 lines of code, and it has been used by more than 41,000 conferences. More than two and a half million users in the scientific community reported using it in 2019.
Circle Hough Transform
The circle Hough Transform (CHT) is a basic feature extraction technique used in digital image processing for detecting circles in imperfect images. The circle candidates are produced by “voting” in the Hough parameter space and then selecting local maxima in an accumulator matrix. It is a specialization of the Hough transform. == Theory == In a two-dimensional space, a circle can be described by: ( x − a ) 2 + ( y − b ) 2 = r 2 ( 1 ) {\displaystyle \left(x-a\right)^{2}+\left(y-b\right)^{2}=r^{2}\ \ \ \ \ (1)} where (a,b) is the center of the circle, and r is the radius. If a 2D point (x,y) is fixed, then the parameters can be found according to (1). The parameter space would be three dimensional, (a, b, r). And all the parameters that satisfy (x, y) would lie on the surface of an inverted right-angled cone whose apex is at (x, y, 0). In the 3D space, the circle parameters can be identified by the intersection of many conic surfaces that are defined by points on the 2D circle. This process can be divided into two stages. The first stage is fixing radius then find the optimal center of circles in a 2D parameter space. The second stage is to find the optimal radius in a one dimensional parameter space. === Find parameters with known radius R === If the radius is fixed, then the parameter space would be reduced to 2D (the position of the circle center). For each point (x, y) on the original circle, it can define a circle centered at (x, y) with radius R according to (1). The intersection point of all such circles in the parameter space would be corresponding to the center point of the original circle. Consider 4 points on a circle in the original image (left). The circle Hough transform is shown in the right. Note that the radius is assumed to be known. For each (x,y) of the four points (white points) in the original image, it can define a circle in the Hough parameter space centered at (x, y) with radius r. An accumulator matrix is used for tracking the intersection point. In the parameter space, the voting number of those points that have a newly defined circle passing through them would be increased by one for every circle. Then the local maxima point (the red point in the center in the right figure) can be found. The position (a, b) of the maxima would be the center of the original circle. === Multiple circles with known radius R === Multiple circles with same radius can be found with the same technique. Note that, in the accumulator matrix (right fig), there would be at least 3 local maxima points. === Accumulator matrix and voting === In practice, an accumulator matrix is introduced to find the intersection point in the parameter space. First, we need to divide the parameter space into “buckets” using a grid and produce an accumulator matrix according to the grid. The element in the accumulator matrix denotes the number of “circles” in the parameter space that are passing through the corresponding grid cell in the parameter space. The number is also called “voting number”. Initially, every element in the matrix is zeros. Then for each “edge” point in the original space, we can formulate a circle in the parameter space and increase the voting number of the grid cell which the circle passes through. This process is called “voting”. After voting, we can find local maxima in the accumulator matrix. The positions of the local maxima are corresponding to the circle centers in the original space. === Find circle parameter with unknown radius === Since the parameter space is 3D, the accumulator matrix would be 3D, too. We can iterate through possible radii; for each radius, we use the previous technique. Finally, find the local maxima in the 3D accumulator matrix. Accumulator array should be A[x,y,r] in the 3D space. Voting should be for each pixels, radius and theta A[x,y,r] += 1 The algorithm : For each A[a,b,r] = 0; Process the filtering algorithm on image Gaussian Blurring, convert the image to grayscale ( grayScaling), make Canny operator, The Canny operator gives the edges on image. Vote on all possible circles in accumulator. The local maximum voted circles of Accumulator A gives the circle Hough space. The maximum voted circle of Accumulator gives the circle. The Incrementing for Best Candidate : For each A[a,b,r] = 0; // fill with zeroes initially, instantiate 3D matrix For each cell(x,y) For each theta t = 0 to 360 // the possible theta 0 to 360 b = y – r sin(t PI / 180); //polar coordinate for center (convert to radians) a = x – r cos(t PI / 180); //polar coordinate for center (convert to radians) A[a,b,r] +=1; //voting end end == Examples == === Find circles in a shoe-print === The original picture (right) is first turned into a binary image (left) using a threshold and Gaussian filter. Then edges (mid) are found from it using canny edge detection. After this, all the edge points are used by the Circle Hough Transform to find underlying circle structure. == Limitations == Since the parameter space of the CHT is three dimensional, it may require lots of storage and computation. Choosing a bigger grid size can ameliorate this problem. However, choosing an appropriate grid size is difficult. Since too coarse a grid can lead to large values of the vote being obtained falsely because many quite different structures correspond to a single bucket. Too fine a grid can lead to structures not being found because votes resulting from tokens that are not exactly aligned end up in different buckets, and no bucket has a large vote. Also, the CHT is not very robust to noise. == Extensions == === Adaptive Hough Transform === J. Illingworth and J. Kittler introduced this method for implementing Hough Transform efficiently. The AHT uses a small accumulator array and the idea of a flexible iterative "coarse to fine" accumulation and search strategy to identify significant peaks in the Hough parameter spaces. This method is substantially superior to the standard Hough Transform implementation in both storage and computational requirements. == Application == === People Counting === Since the head would be similar to a circle in an image, CHT can be used for detecting heads in a picture, so as to count the number of persons in the image. === Brain Aneurysm Detection === Modified Hough Circle Transform (MHCT) is used on the image extracted from Digital Subtraction Angiogram (DSA) to detect and classify aneurysms type. == Implementation code == Circle Detection via Standard Hough Transform, by Amin Sarafraz, Mathworks (File Exchange) Hough Circle Transform, OpenCV-Python Tutorials (archived version on archive.org)
Collateral freedom
Collateral freedom is an anti-censorship strategy that attempts to make it economically prohibitive for censors to block content on the Internet. This is achieved by hosting content on cloud services that are considered by censors to be "too important to block", and then using encryption to prevent censors from identifying requests for censored information that is hosted among other content, forcing censors to either allow access to the censored information or take down entire services.
Color vision
Color vision (CV), a feature of visual perception, is an ability to perceive differences between light composed of different frequencies independently of light intensity. Color perception is a part of the larger visual system and is mediated by a complex process between neurons that begins with differential stimulation of different types of photoreceptors by light entering the eye. Those photoreceptors then emit outputs that are propagated through many layers of neurons ultimately leading to higher cognitive functions in the brain. Color vision is found in many animals and is mediated by similar underlying mechanisms with common types of biological molecules and a complex history of the evolution of color vision within different animal taxa. In primates, color vision may have evolved under selective pressure for a variety of visual tasks including the foraging for nutritious young leaves, ripe fruit, and flowers, as well as detecting predator camouflage and emotional states in other primates. == Wavelength == Isaac Newton discovered that white light after being split into its component colors when passed through a dispersive prism could be recombined to make white light by passing them through a different prism. The visible light spectrum ranges from about 380 to 740 nanometers. Spectral colors (colors that are produced by a narrow band of wavelengths) such as red, orange, yellow, green, cyan, blue, and violet can be found in this range. These spectral colors do not refer to a single wavelength, but rather to a set of wavelengths: red, 625–740 nm; orange, 590–625 nm; yellow, 565–590 nm; green, 500–565 nm; cyan, 485–500 nm; blue, 450–485 nm; violet, 380–450 nm. Wavelengths longer or shorter than this range are called infrared or ultraviolet, respectively. Humans cannot generally see these wavelengths, but other animals may. === Hue detection === Sufficient differences in wavelength cause a difference in the perceived hue; the just-noticeable difference in wavelength varies from about 1 nm in the blue-green and yellow wavelengths to 10 nm and more in the longer red and shorter blue wavelengths. Although the human eye can distinguish up to a few hundred hues, when those pure spectral colors are mixed together or diluted with white light, the number of distinguishable chromaticities can be much higher. In very low light levels, vision is scotopic: light is detected by rod cells of the retina. Rods are maximally sensitive to wavelengths near 500 nm and play little, if any, role in color vision. In brighter light, such as daylight, vision is photopic: light is detected by cone cells which are responsible for color vision. Cones are sensitive to a range of wavelengths, but are most sensitive to wavelengths near 555 nm. Between these regions, mesopic vision comes into play and both rods and cones provide signals to the retinal ganglion cells. The shift in color perception from dim light to daylight gives rise to differences known as the Purkinje effect. The perception of "white" is formed by the entire spectrum of visible light, or by mixing colors of just a few wavelengths in animals with few types of color receptors. In humans, white light can be perceived by combining wavelengths such as red, green, and blue, or just a pair of complementary colors such as blue and yellow. === Non-spectral colors === There are a variety of colors in addition to spectral colors and their hues. These include grayscale colors, shades of colors obtained by mixing grayscale colors with spectral colors, violet-red colors, impossible colors, and metallic colors. Grayscale colors include white, gray, and black. Rods contain rhodopsin, which reacts to light intensity, providing grayscale coloring. Shades include colors such as pink or brown. Pink is obtained from mixing red and white. Brown may be obtained from mixing orange with gray or black. Navy is obtained from mixing blue and black. Violet-red colors include hues and shades of magenta. The light spectrum is a line on which violet is one end and the other is red, and yet we see hues of purple that connect those two colors. Impossible colors are a combination of cone responses that cannot be naturally produced. For example, medium cones cannot be activated completely on their own; if they were, we would see a 'hyper-green' color. == Dimensionality == Color vision is categorized foremost according to the dimensionality of the color gamut, which is defined by the number of primaries required to represent the color vision. This is generally equal to the number of photopsins expressed: a correlation that holds for vertebrates but not invertebrates. The common vertebrate ancestor possessed four photopsins (expressed in cones) plus rhodopsin (expressed in rods), so was tetrachromatic. However, many vertebrate lineages have lost one or many photopsin genes, leading to lower-dimension color vision. The dimensions of color vision range from 1-dimensional and up: == Physiology of color perception == Perception of color begins with specialized retinal cells known as cone cells. Cone cells contain different forms of opsin – a pigment protein – that have different spectral sensitivities. Humans contain three types, resulting in trichromatic color vision. Each individual cone contains pigments composed of opsin apoprotein covalently linked to a light-absorbing prosthetic group: either 11-cis-hydroretinal or, more rarely, 11-cis-dehydroretinal. The cones are conventionally labeled according to the ordering of the wavelengths of the peaks of their spectral sensitivities: short (S), medium (M), and long (L) cone types. These three types do not correspond well to particular colors as we know them. Rather, the perception of color is achieved by a complex process that starts with the differential output of these cells in the retina and which is finalized in the visual cortex and associative areas of the brain. For example, while the L cones have been referred to simply as red receptors, microspectrophotometry has shown that their peak sensitivity is in the greenish-yellow region of the spectrum. Similarly, the S cones and M cones do not directly correspond to blue and green, although they are often described as such. The RGB color model, therefore, is a convenient means for representing color but is not directly based on the types of cones in the human eye. The peak response of human cone cells varies, even among individuals with typical color vision; in some non-human species this polymorphic variation is even greater, and it may well be adaptive. === Theories === Two complementary theories of color vision are the trichromatic theory and the opponent process theory. The trichromatic theory, or Young–Helmholtz theory, proposed in the 19th century by Thomas Young and Hermann von Helmholtz, posits three types of cones preferentially sensitive to blue, green, and red, respectively. Others have suggested that the trichromatic theory is not specifically a theory of color vision but a theory of receptors for all vision, including color but not specific or limited to it. Equally, it has been suggested that the relationship between the phenomenal opponency described by Ewald Hering and the physiological opponent processes are not straightforward (see below), making of physiological opponency a mechanism that is relevant to the whole of vision, and not just to color vision alone. Hering proposed the opponent process theory in 1872. It states that the visual system interprets color in an antagonistic way: red vs. green, blue vs. yellow, black vs. white. Both theories are generally accepted as valid, describing different stages in visual physiology, visualized in the adjacent diagram. Green–magenta and blue–yellow are scales with mutually exclusive boundaries. In the same way that there cannot exist a "slightly negative" positive number, a single eye cannot perceive a bluish-yellow or a reddish-green. Although these two theories are both currently widely accepted theories, past and more recent work has led to criticism of the opponent process theory, stemming from a number of what are presented as discrepancies in the standard opponent process theory. For example, the phenomenon of an after-image of complementary color can be induced by fatiguing the cells responsible for color perception, by staring at a vibrant color for a length of time, and then looking at a white surface. This phenomenon of complementary colors shows that cyan, rather than green, is the complement of red, and that magenta, rather than red, is the complement of green. It therefore also shows that the reddish-green color supposed to be impossible by opponent process theory is actually the color yellow. Although this phenomenon is more readily explained by the trichromatic theory, explanations for the discrepancy may include alterations to the opponent process theory, such as redefining the opponent colors as red vs. cyan, to reflect this effect. Despite such criticis