Contrast set learning is a form of association rule learning that seeks to identify meaningful differences between separate groups by reverse-engineering the key predictors that identify for each particular group. For example, given a set of attributes for a pool of students (labeled by degree type), a contrast set learner would identify the contrasting features between students seeking bachelor's degrees and those working toward PhD degrees. == Overview == A common practice in data mining is to classify, to look at the attributes of an object or situation and make a guess at what category the observed item belongs to. As new evidence is examined (typically by feeding a training set to a learning algorithm), these guesses are refined and improved. Contrast set learning works in the opposite direction. While classifiers read a collection of data and collect information that is used to place new data into a series of discrete categories, contrast set learning takes the category that an item belongs to and attempts to reverse engineer the statistical evidence that identifies an item as a member of a class. That is, contrast set learners seek rules associating attribute values with changes to the class distribution. They seek to identify the key predictors that contrast one classification from another. For example, an aerospace engineer might record data on test launches of a new rocket. Measurements would be taken at regular intervals throughout the launch, noting factors such as the trajectory of the rocket, operating temperatures, external pressures, and so on. If the rocket launch fails after a number of successful tests, the engineer could use contrast set learning to distinguish between the successful and failed tests. A contrast set learner will produce a set of association rules that, when applied, will indicate the key predictors of each failed tests versus the successful ones (the temperature was too high, the wind pressure was too high, etc.). Contrast set learning is a form of association rule learning. Association rule learners typically offer rules linking attributes commonly occurring together in a training set (for instance, people who are enrolled in four-year programs and take a full course load tend to also live near campus). Instead of finding rules that describe the current situation, contrast set learners seek rules that differ meaningfully in their distribution across groups (and thus, can be used as predictors for those groups). For example, a contrast set learner could ask, “What are the key identifiers of a person with a bachelor's degree or a person with a PhD, and how do people with PhD's and bachelor’s degrees differ?” Standard classifier algorithms, such as C4.5, have no concept of class importance (that is, they do not know if a class is "good" or "bad"). Such learners cannot bias or filter their predictions towards certain desired classes. As the goal of contrast set learning is to discover meaningful differences between groups, it is useful to be able to target the learned rules towards certain classifications. Several contrast set learners, such as MINWAL or the family of TAR algorithms, assign weights to each class in order to focus the learned theories toward outcomes that are of interest to a particular audience. Thus, contrast set learning can be thought of as a form of weighted class learning. === Example: Supermarket Purchases === The differences between standard classification, association rule learning, and contrast set learning can be illustrated with a simple supermarket metaphor. In the following small dataset, each row is a supermarket transaction and each "1" indicates that the item was purchased (a "0" indicates that the item was not purchased): Given this data, Association rule learning may discover that customers that buy onions and potatoes together are likely to also purchase hamburger meat. Classification may discover that customers that bought onions, potatoes, and hamburger meats were purchasing items for a cookout. Contrast set learning may discover that the major difference between customers shopping for a cookout and those shopping for an anniversary dinner are that customers acquiring items for a cookout purchase onions, potatoes, and hamburger meat (and do not purchase foie gras or champagne). == Treatment learning == Treatment learning is a form of weighted contrast-set learning that takes a single desirable group and contrasts it against the remaining undesirable groups (the level of desirability is represented by weighted classes). The resulting "treatment" suggests a set of rules that, when applied, will lead to the desired outcome. Treatment learning differs from standard contrast set learning through the following constraints: Rather than seeking the differences between all groups, treatment learning specifies a particular group to focus on, applies a weight to this desired grouping, and lumps the remaining groups into one "undesired" category. Treatment learning has a stated focus on minimal theories. In practice, treatment are limited to a maximum of four constraints (i.e., rather than stating all of the reasons that a rocket differs from a skateboard, a treatment learner will state one to four major differences that predict for rockets at a high level of statistical significance). This focus on simplicity is an important goal for treatment learners. Treatment learning seeks the smallest change that has the greatest impact on the class distribution. Conceptually, treatment learners explore all possible subsets of the range of values for all attributes. Such a search is often infeasible in practice, so treatment learning often focuses instead on quickly pruning and ignoring attribute ranges that, when applied, lead to a class distribution where the desired class is in the minority. === Example: Boston housing data === The following example demonstrates the output of the treatment learner TAR3 on a dataset of housing data from the city of Boston (a nontrivial public dataset with over 500 examples). In this dataset, a number of factors are collected for each house, and each house is classified according to its quality (low, medium-low, medium-high, and high). The desired class is set to "high", and all other classes are lumped together as undesirable. The output of the treatment learner is as follows: Baseline class distribution: low: 29% medlow: 29% medhigh: 21% high: 21% Suggested Treatment: [PTRATIO=[12.6..16), RM=[6.7..9.78)] New class distribution: low: 0% medlow: 0% medhigh: 3% high: 97% With no applied treatments (rules), the desired class represents only 21% of the class distribution. However, if one filters the data set for houses with 6.7 to 9.78 rooms and a neighborhood parent-teacher ratio of 12.6 to 16, then 97% of the remaining examples fall into the desired class (high-quality houses). == Algorithms == There are a number of algorithms that perform contrast set learning. The following subsections describe two examples. === STUCCO === The STUCCO contrast set learner treats the task of learning from contrast sets as a tree search problem where the root node of the tree is an empty contrast set. Children are added by specializing the set with additional items picked through a canonical ordering of attributes (to avoid visiting the same nodes twice). Children are formed by appending terms that follow all existing terms in a given ordering. The formed tree is searched in a breadth-first manner. Given the nodes at each level, the dataset is scanned and the support is counted for each group. Each node is then examined to determine if it is significant and large, if it should be pruned, and if new children should be generated. After all significant contrast sets are located, a post-processor selects a subset to show to the user - the low order, simpler results are shown first, followed by the higher order results which are "surprising and significantly different." The support calculation comes from testing a null hypothesis that the contrast set support is equal across all groups (i.e., that contrast set support is independent of group membership). The support count for each group is a frequency value that can be analyzed in a contingency table where each row represents the truth value of the contrast set and each column variable indicates the group membership frequency. If there is a difference in proportions between the contrast set frequencies and those of the null hypothesis, the algorithm must then determine if the differences in proportions represent a relation between variables or if it can be attributed to random causes. This can be determined through a chi-square test comparing the observed frequency count to the expected count. Nodes are pruned from the tree when all specializations of the node can never lead to a significant and large contrast set. The decision to prune is based on: The minimum deviation size: The maximum difference between the support
Artificial intelligence controversies
The controversies surrounding artificial intelligence encompass a broad range of public, academic, and political debates regarding the societal effects of artificial intelligence (AI). These debates intensified particularly in the late 2010s and 2020s, coinciding with an accelerated period of development known as the AI boom. While advocates emphasize the technology's potential to solve complex problems and enhance human quality of life, detractors highlight a wide array of dangers and challenges. These include concerns over ethics, plagiarism and theft, fraud, safety and alignment, environmental impacts, technological unemployment, and the spread of misinformation. It also covers severe future or theoretical challenges, such as the emergence of artificial superintelligence and existential risks. == 2016 == === Microsoft Tay chatbot (2016) === On March 23, 2016, Microsoft released Tay, a chatbot designed to mimic the language patterns of a 19-year-old American girl and learn from interactions with Twitter users. Soon after its launch, Tay began posting racist, sexist, and otherwise inflammatory tweets after Twitter users deliberately taught it offensive phrases and exploited its "repeat after me" capability. Examples of controversial outputs included Holocaust denial and calls for genocide using racial slurs. Within 16 hours of its release, Microsoft suspended the Twitter account, deleted the offensive tweets, and stated that Tay had suffered from a "coordinated attack by a subset of people" that "exploited a vulnerability." Tay was briefly and accidentally re-released on March 30 during testing, after which it was permanently shut down. Microsoft CEO Satya Nadella later stated that Tay "has had a great influence on how Microsoft is approaching AI" and taught the company the importance of taking accountability. == 2022 == === Voiceverse NFT plagiarism scandal (2022) === On January 14, 2022, voice actor Troy Baker announced a partnership with Voiceverse, a blockchain-based company that marketed proprietary AI voice cloning technology as non-fungible tokens (NFT), triggering immediate backlash over environmental concerns, fears that AI could displace human voice actors, and concerns about fraud. Later that same day, the pseudonymous creator of 15.ai—a free, non-commercial AI voice synthesis research project—revealed through server logs that Voiceverse had used 15.ai to generate voice samples, pitch-shifted them to make them unrecognizable, and falsely marketed them as their own proprietary technology before selling them as NFTs; the developer of 15.ai had previously stated that they had no interest in incorporating NFTs into their work. Voiceverse confessed within an hour and stated that their marketing team had used 15.ai without attribution while rushing to create a demo. News publications and AI watchdog groups universally characterized the incident as theft stemming from generative artificial intelligence. === Théâtre D'opéra Spatial (2022) === On August 29, 2022, Jason Michael Allen won first place in the "emerging artist" (non-professional) division of the "Digital Arts/Digitally-Manipulated Photography" category of the Colorado State Fair's fine arts competition with Théâtre D'opéra Spatial, a digital artwork created using the AI image generator Midjourney, Adobe Photoshop, and AI upscaling tools, becoming one of the first images made using generative AI to win such a prize. Allen disclosed his use of Midjourney when submitting, though the judges did not know it was an AI tool but stated they would have awarded him first place regardless. While there was little contention about the image at the fair, reactions to the win on social media were negative. On September 5, 2023, the United States Copyright Office ruled that the work was not eligible for copyright protection as the human creative input was de minimis and that copyright rules "exclude works produced by non-humans." == 2023 == === Statements on AI risk (2023) === On March 22, 2023, the Future of Life Institute published an open letter calling on "all AI labs to immediately pause for at least 6 months the training of AI systems more powerful than GPT-4", citing risks such as AI-generated propaganda, extreme automation of jobs, human obsolescence, and a society-wide loss of control. The letter, published a week after the release of OpenAI's GPT-4, asserted that current large language models were "becoming human-competitive at general tasks". It received more than 30,000 signatures, including academic AI researchers and industry CEOs such as Yoshua Bengio, Stuart Russell, Elon Musk, Steve Wozniak and Yuval Noah Harari. The letter was criticized for diverting attention from more immediate societal risks such as algorithmic biases, with Timnit Gebru and others arguing that it amplified "some futuristic, dystopian sci-fi scenario" instead of current problems with AI. On May 30, 2023, the Center for AI Safety released a one-sentence statement signed by hundreds of artificial intelligence experts and other notable figures: "Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war." Signatories included Turing laureates Geoffrey Hinton and Yoshua Bengio, as well as the scientific and executive leaders of several major AI companies, including Sam Altman, Demis Hassabis, and Bill Gates. The statement prompted responses from political leaders, including UK Prime Minister Rishi Sunak, who retweeted it with a statement that the UK government would look carefully into it, and White House Press Secretary Karine Jean-Pierre, who commented that AI "is one of the most powerful technologies that we see currently in our time." Skeptics, including from Human Rights Watch, argued that scientists should focus on known risks of AI instead of speculative future risks. === Removal of Sam Altman from OpenAI (2023) === On November 17, 2023, OpenAI's board of directors ousted co-founder and chief executive Sam Altman, stating that "the board no longer has confidence in his ability to continue leading OpenAI." The removal was precipitated by employee concerns about his handling of artificial intelligence safety and allegations of abusive behavior. Altman was reinstated on November 22 after pressure from employees and investors, including a letter signed by 745 of OpenAI's 770 employees threatening mass resignations if the board did not resign. The removal and subsequent reinstatement caused widespread reactions, including Microsoft's stock falling nearly three percent following the initial announcement and then rising over two percent to an all-time high after Altman was hired to lead a Microsoft AI research team before his reinstatement. The incident also prompted investigations from the Competition and Markets Authority and the Federal Trade Commission into Microsoft's relationship with OpenAI. == 2024 == === Taylor Swift deepfake pornography controversy (2024) === In late January 2024, sexually explicit AI-generated deepfake images of Taylor Swift were proliferated on X, with one post reported to have been seen over 47 million times before its removal. Disinformation research firm Graphika traced the images back to 4chan, while members of a Telegram group had discussed ways to circumvent censorship safeguards of AI image generators to create pornographic images of celebrities. The images prompted responses from anti-sexual assault advocacy groups, US politicians, and Swifties. Microsoft CEO Satya Nadella called the incident "alarming and terrible." X briefly blocked searches of Swift's name on January 27, 2024, and Microsoft enhanced its text-to-image model safeguards to prevent future abuse. On January 30, US senators Dick Durbin, Lindsey Graham, Amy Klobuchar, and Josh Hawley introduced a bipartisan bill that would allow victims to sue individuals who produced or possessed "digital forgeries" with intent to distribute, or those who received the material knowing it was made without consent. === Google Gemini image generation controversy (2024) === In February 2024, social media users reported that Google's Gemini chatbot was generating images that featured people of color and women in historically inaccurate contexts—such as Vikings, Nazi soldiers, and the Founding Fathers—and refusing prompts to generate images of white people. The images were derided on social media, including by conservatives who cited them as evidence of Google's "wokeness", and criticized by Elon Musk, who denounced Google's products as biased and racist. In response, Google paused Gemini's ability to generate images of people. Google executive Prabhakar Raghavan released a statement explaining that Gemini had "overcompensate[d]" in its efforts to strive for diversity and acknowledging that the images were "embarrassing and wrong". Google CEO Sundar Pichai called the incident offensive and unacceptable in an internal memo, promising struc
AI Pair Programmers: Free vs Paid (2026)
Trying to pick the best AI pair programmer? An AI pair programmer is software that uses machine learning to help you get more done — it scales effortlessly from a single task to thousands. The best picks balance beginner-friendly simplicity with the depth power users need, and they ship updates often. Whether you are a beginner or a pro, the right AI pair programmer slots into your workflow and pays for itself fast. This guide breaks down the top picks, their pros and cons, and who each one is best for.
Ali Farhadi
Ali Farhadi is a professor of computer science and currently serves as the CEO of the Allen Institute for Artificial Intelligence (AI2). == Career == Farhadi is an AI professor in the Paul G. Allen School of Computer Science & Engineering at the University of Washington in Seattle. == Recognitions == Among other awards and recognitions, Farhadi was one of the winners of the 2017 Sloan Research Fellowship granted by the Alfred P. Sloan Foundation. == Personal life == Farhadi is married to Hanna Hajishirzi, a computer science professor at the University of Washington and senior director at the Allen Institute who studies natural language processing.
Sumio Watanabe
Sumio Watanabe (渡辺 澄夫, Watanabe Sumio; born 1959) is a Japanese mathematician and engineer working in probability theory, applied algebraic geometry and Bayesian statistics. He is currently a professor at Tokyo Institute of Technology in the Department of Computational Intelligence and Systems Science. He is the author of the text, Algebraic Geometry and Statistical Learning Theory, which proposes a generalization of Fisher's regular statistical theory to singular statistical models. == Books == Mathematical Theory of Bayesian Statistics, CRC Press, 2018, ISBN 9781482238068 Algebraic Geometry and Statistical Learning Theory, Cambridge University Press, 2009.
Apache CarbonData
Apache CarbonData is a free and open-source column-oriented data storage format of the Apache Hadoop ecosystem. It is similar to the other columnar-storage file formats available in Hadoop namely RCFile and ORC. It is compatible with most of the data processing frameworks in the Hadoop environment. It provides efficient data compression and encoding schemes with enhanced performance to handle complex data in bulk. == History == CarbonData was developed at Huawei in 2013. The project was donated to the Apache Community in 2015 submitted to the Apache Incubator in June 2016. The project won top honors in the BlackDuck 2016 Open Source Rookies of the Year's Big Data category. Apache CarbonData has been a top-level Apache Software Foundation (ASF)-sponsored project since May 1, 2017.
Unsupervised learning
Unsupervised learning is a framework in machine learning where, in contrast to supervised learning, algorithms learn patterns exclusively from unlabeled data. Other frameworks in the spectrum of supervisions include weak- or semi-supervision, where a small portion of the data is tagged, and self-supervision. Some researchers consider self-supervised learning a form of unsupervised learning. Conceptually, unsupervised learning divides into the aspects of data, training, algorithm, and downstream applications. Typically, the dataset is harvested cheaply "in the wild", such as massive text corpus obtained by web crawling, with only minor filtering (such as Common Crawl). This compares favorably to supervised learning, where the dataset (such as the ImageNet1000) is typically constructed manually, which is much more expensive. There are algorithms designed specifically for unsupervised learning, such as clustering algorithms like k-means, dimensionality reduction techniques like principal component analysis (PCA), Boltzmann machine learning, and autoencoders. After the rise of deep learning, most large-scale unsupervised learning has been done by training general-purpose neural network architectures by gradient descent, adapted to performing unsupervised learning by designing an appropriate training procedure. Sometimes a trained model can be used as-is, but more often they are modified for downstream applications. For example, the generative pretraining method trains a model to generate a textual dataset, before finetuning it for other applications, such as text classification. As another example, autoencoders are trained to produce good features, which can then be used as a module for other models, such as in a latent diffusion model. == Tasks == Tasks are often categorized as discriminative (recognition) or generative (imagination). Often but not always, discriminative tasks use supervised methods and generative tasks use unsupervised (see Venn diagram); however, the separation is very hazy. For example, object recognition favors supervised learning but unsupervised learning can also cluster objects into groups. Furthermore, as progress marches onward, some tasks employ both methods, and some tasks swing from one to another. For example, image recognition started off as heavily supervised, but became hybrid by employing unsupervised pre-training, and then moved towards supervision again with the advent of dropout, ReLU, and adaptive learning rates. A typical generative task is as follows. At each step, a datapoint is sampled from the dataset, and part of the data is removed, and the model must infer the removed part. This is particularly clear for the denoising autoencoders and BERT. == Neural network architectures == === Training === During the learning phase, an unsupervised network tries to mimic the data it is given and uses the error in its mimicked output to correct itself (i.e. correct its weights and biases). Sometimes the error is expressed as a low probability that the erroneous output occurs, or it might be expressed as an unstable high energy state in the network. In contrast to supervised methods' dominant use of backpropagation, unsupervised learning also employs other methods including: Hopfield learning rule, Boltzmann learning rule, Contrastive Divergence, Wake Sleep, Variational Inference, Maximum Likelihood, Maximum A Posteriori, Gibbs Sampling, and backpropagating reconstruction errors or hidden state reparameterizations. See the table below for more details. === Energy === An energy function is a macroscopic measure of a network's activation state. In Boltzmann machines, it plays the role of the Cost function. This analogy with physics is inspired by Ludwig Boltzmann's analysis of a gas' macroscopic energy from the microscopic probabilities of particle motion p ∝ e − E / k T {\displaystyle p\propto e^{-E/kT}} , where k is the Boltzmann constant and T is temperature. In the RBM network the relation is p = e − E / Z {\displaystyle p=e^{-E}/Z} , where p {\displaystyle p} and E {\displaystyle E} vary over every possible activation pattern and Z = ∑ All Patterns e − E ( pattern ) {\displaystyle \textstyle {Z=\sum _{\scriptscriptstyle {\text{All Patterns}}}e^{-E({\text{pattern}})}}} . To be more precise, p ( a ) = e − E ( a ) / Z {\displaystyle p(a)=e^{-E(a)}/Z} , where a {\displaystyle a} is an activation pattern of all neurons (visible and hidden). Hence, some early neural networks bear the name Boltzmann Machine. Paul Smolensky calls − E {\displaystyle -E\,} the Harmony. A network seeks low energy which is high Harmony. === Networks === This table shows connection diagrams of various unsupervised networks, the details of which will be given in the section Comparison of Networks. Circles are neurons and edges between them are connection weights. As network design changes, features are added on to enable new capabilities or removed to make learning faster. For instance, neurons change between deterministic (Hopfield) and stochastic (Boltzmann) to allow robust output, weights are removed within a layer (RBM) to hasten learning, or connections are allowed to become asymmetric (Helmholtz). Of the networks bearing people's names, only Hopfield worked directly with neural networks. Boltzmann and Helmholtz came before artificial neural networks, but their work in physics and physiology inspired the analytical methods that were used. === History === === Specific Networks === Here, we highlight some characteristics of select networks. The details of each are given in the comparison table below. Hopfield Network Ferromagnetism inspired Hopfield networks. A neuron corresponds to an iron domain with binary magnetic moments Up and Down, and neural connections correspond to the domain's influence on each other. Symmetric connections enable a global energy formulation. During inference the network updates each state using the standard activation step function. Symmetric weights and the right energy functions guarantees convergence to a stable activation pattern. Asymmetric weights are difficult to analyze. Hopfield nets are used as Content Addressable Memories (CAM). Boltzmann Machine These are stochastic Hopfield nets. Their state value is sampled from this pdf as follows: suppose a binary neuron fires with the Bernoulli probability p(1) = 1/3 and rests with p(0) = 2/3. One samples from it by taking a uniformly distributed random number y, and plugging it into the inverted cumulative distribution function, which in this case is the step function thresholded at 2/3. The inverse function = { 0 if x <= 2/3, 1 if x > 2/3 }. Sigmoid Belief Net Introduced by Radford Neal in 1992, this network applies ideas from probabilistic graphical models to neural networks. A key difference is that nodes in graphical models have pre-assigned meanings, whereas Belief Net neurons' features are determined after training. The network is a sparsely connected directed acyclic graph composed of binary stochastic neurons. The learning rule comes from Maximum Likelihood on p(X): Δwij ∝ {\displaystyle \propto } sj (si - pi), where pi = 1 / ( 1 + eweighted inputs into neuron i ). sj's are activations from an unbiased sample of the posterior distribution and this is problematic due to the Explaining Away problem raised by Judea Perl. Variational Bayesian methods uses a surrogate posterior and blatantly disregard this complexity. Deep Belief Network Introduced by Hinton, this network is a hybrid of RBM and Sigmoid Belief Network. The top 2 layers is an RBM and the second layer downwards form a sigmoid belief network. One trains it by the stacked RBM method and then throw away the recognition weights below the top RBM. As of 2009, 3-4 layers seems to be the optimal depth. Helmholtz machine These are early inspirations for the Variational Auto Encoders. Its 2 networks combined into one—forward weights operates recognition and backward weights implements imagination. It is perhaps the first network to do both. Helmholtz did not work in machine learning but he inspired the view of "statistical inference engine whose function is to infer probable causes of sensory input". the stochastic binary neuron outputs a probability that its state is 0 or 1. The data input is normally not considered a layer, but in the Helmholtz machine generation mode, the data layer receives input from the middle layer and has separate weights for this purpose, so it is considered a layer. Hence this network has 3 layers. Variational autoencoder These are inspired by Helmholtz machines and combines probability network with neural networks. An Autoencoder is a 3-layer CAM network, where the middle layer is supposed to be some internal representation of input patterns. The encoder neural network is a probability distribution qφ(z given x) and the decoder network is pθ(x given z). The weights are named phi & theta rather than W and V as in Helmholtz—a cosmetic difference. These 2 networks h