An Artificial Intelligence Glossary
Artificial intelligence is much larger than large language models. This reference spans symbolic AI, machine learning, neural networks, computer vision, reinforcement learning, autonomy, robotics, causal reasoning, generative systems, safety, landmark researchers, and the projects that changed the field.
Language models are one important branch of a much older and broader field that includes perception, control, search, planning, probability, causal reasoning, robotics, autonomy, and symbolic systems.
The vocabulary of AI makes more sense when the ideas are connected to the people, experiments, datasets, systems, and research programs that made them concrete.
A* search
A* is a best-first search algorithm that expands paths according to the cost already incurred plus a heuristic estimate of the cost remaining. With an admissible heuristic, A* can find an optimal path while avoiding much of the search performed by uninformed methods.
Why it matters: A* remains a foundational example of how AI can combine exact computation with problem-specific knowledge to search large spaces efficiently.
Active learning
Active learning is a training strategy in which a model or learning system selects the examples for which labels would be most informative, rather than accepting a fixed labeled dataset.
Why it matters: It is useful when labels are expensive, expert time is scarce, or the system can improve faster by choosing what humans should annotate next.
Actor-critic
Actor-critic methods divide reinforcement learning into two interacting components: an actor that selects actions and a critic that estimates how good states or actions are.
Why it matters: The pattern underlies many modern policy-gradient systems because it combines direct policy learning with value estimation.
Adversarial example
An adversarial example is an input intentionally or accidentally modified in a way that can cause a machine-learning model to make a surprising error despite the change appearing small or irrelevant to a person.
Why it matters: Adversarial examples demonstrate that high benchmark accuracy does not guarantee robust behavior under unusual or strategically chosen inputs.
AI agent
An AI agent is a system that observes an environment, maintains or derives state, selects actions, and acts in pursuit of an objective. Agents may operate in software environments, physical environments, or both.
Why it matters: The agent framing shifts attention from one-shot prediction to sequences of decisions, feedback, memory, planning, and consequences.
AI alignment
AI alignment is the research and engineering problem of making AI systems behave in ways that are consistent with intended goals, constraints, and human values.
Why it matters: Alignment becomes more difficult as systems gain autonomy, generality, and the ability to pursue multi-step objectives in environments that designers cannot fully specify.
Alan Turing
Alan Turing was a mathematician and computer scientist whose 1950 paper Computing Machinery and Intelligence asked whether machines can think and proposed the imitation game later known as the Turing test.
Why it matters: Turing's paper remains foundational because it shifted debate toward observable intelligent behavior and anticipated many enduring questions about learning machines.
AlexNet
AlexNet is the deep convolutional neural network by Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton that won the 2012 ImageNet competition by a large margin using GPU-trained deep learning.
Why it matters: AlexNet is widely treated as a turning point in modern computer vision because it demonstrated the practical combination of large labeled datasets, GPUs, deep networks, and improved training techniques.
Algorithmic auditing
Algorithmic auditing is the systematic evaluation of an automated system's behavior, performance, data, controls, or impacts, often with attention to subgroups and deployment context.
Why it matters: Auditing provides a way to move from abstract claims of fairness or reliability to empirical testing of how systems actually behave.
Algorithmic bias
Algorithmic bias is systematic variation in an automated system's behavior or outcomes that disadvantages or misrepresents groups, often because of training data, measurement choices, labels, objectives, deployment context, or product design.
Why it matters: Bias can enter an AI system long before model training and can persist even when aggregate accuracy looks strong.
Allen Newell
Allen Newell was a computer scientist who, with Herbert A. Simon and collaborators, helped create early symbolic AI systems including Logic Theorist and General Problem Solver.
Why it matters: His work shaped research on search, problem solving, cognitive architectures, and the idea that intelligence could be studied through computational symbol systems.
Alpha-beta pruning
Alpha-beta pruning is an optimization for minimax game-tree search that eliminates branches that cannot affect the final decision under the current bounds.
Why it matters: It made deeper game-tree search practical and remains a classic example of using mathematical bounds to reduce computation without changing the result.
AlphaFold
AlphaFold is a DeepMind project that uses machine learning to predict protein structure from amino-acid sequence. AlphaFold2's 2021 Nature paper marked a major advance in protein-structure prediction.
Why it matters: AlphaFold is an important example of AI producing scientific value outside language and consumer software, and of learned representations being combined with domain structure.
AlphaGo
AlphaGo was a DeepMind system that combined deep neural networks, reinforcement learning, and tree search to play the game of Go. In 2016, Nature published the system that defeated professional player Fan Hui, followed by the widely watched match against Lee Sedol.
Why it matters: AlphaGo demonstrated the power of combining learned policy and value functions with search, and helped accelerate research in deep reinforcement learning and self-play.
Andrew Barto
Andrew G. Barto is a computer scientist whose work with Richard Sutton helped establish the modern framework and vocabulary of reinforcement learning, including temporal-difference methods and the influential textbook Reinforcement Learning: An Introduction.
Why it matters: Barto's work connects machine learning with ideas from control, psychology, and neuroscience and helped make reinforcement learning a coherent field.
Arthur Samuel
Arthur Samuel was an IBM researcher known for pioneering computer checkers and for early work in machine learning, including programs that improved through experience.
Why it matters: His checkers work is an important early example of a system learning evaluation and strategy rather than relying entirely on fixed rules.
Artificial general intelligence
Artificial general intelligence, or AGI, is a loosely defined term for AI with broad, flexible competence across many intellectual tasks rather than strong performance in one narrow domain.
Why it matters: There is no universally accepted technical test for AGI, so claims about whether a system is or is not AGI depend heavily on the definition being used.
Artificial intelligence
Artificial intelligence is the broad field concerned with building machines that perform tasks associated with perception, reasoning, learning, language, planning, decision-making, or intelligent action.
Why it matters: AI includes symbolic reasoning, machine learning, computer vision, robotics, reinforcement learning, probabilistic methods, search, planning, and many other approaches. Large language models are one branch of the field, not the field itself.
Artificial narrow intelligence
Artificial narrow intelligence describes systems that perform well within a bounded task or domain, such as image classification, speech recognition, route planning, or industrial inspection.
Why it matters: Most deployed AI systems are narrow in the sense that their competence depends on particular tasks, data, interfaces, and operating assumptions.
Attention
Attention is a mechanism that computes how strongly one representation should use information from other representations when producing a new representation. Different forms of attention appear in language, vision, audio, and multimodal models.
Why it matters: Attention lets models dynamically focus computation on relevant parts of an input and is central to Transformer architectures.
Automation bias
Automation bias is the tendency for people to over-rely on automated recommendations or fail to seek contradictory information when a system appears authoritative.
Why it matters: Human oversight can fail if users are trained to defer to the tool, lack domain context, or are under time pressure.
Autonomous vehicle
An autonomous vehicle is a vehicle that uses sensing, localization, perception, prediction, planning, and control to perform some or all of the driving task without continuous human control.
Why it matters: Autonomous driving is a systems problem that combines computer vision, sensor fusion, robotics, maps, planning, machine learning, verification, and real-time safety engineering.
Autonomy
Autonomy is the capacity of a system to perceive conditions, make decisions, and act with reduced direct human control over some period of time or scope of activity.
Why it matters: Autonomy is a spectrum. A system can automate a narrow function, execute a bounded mission, or adapt its behavior across changing conditions.
Backpropagation
Backpropagation is an efficient method for computing how changes in neural-network parameters affect a loss function by propagating error gradients backward through the computational graph.
Why it matters: The 1986 paper by David Rumelhart, Geoffrey Hinton, and Ronald Williams helped popularize backpropagation for learning useful internal representations in multilayer networks.
Bayesian network
A Bayesian network is a directed acyclic graph that represents probabilistic relationships among variables and supports reasoning under uncertainty by factorizing a joint probability distribution.
Why it matters: Bayesian networks made it practical to combine structured domain knowledge with probabilistic inference and are closely associated with Judea Pearl's work.
Benchmark
A benchmark is a standardized dataset, task, environment, or suite used to compare AI systems under a shared evaluation procedure.
Why it matters: Benchmarks make progress measurable, but performance can become misleading when the benchmark is narrow, contaminated, saturated, or poorly matched to deployment.
Bias-variance tradeoff
The bias-variance tradeoff describes the tension between models that are too simple to capture the underlying pattern and models that are so flexible that they fit noise or peculiarities of the training data.
Why it matters: It remains a useful conceptual tool for understanding generalization, overfitting, regularization, and model complexity.
Calibration
Calibration describes how well a model's stated confidence corresponds to the frequency with which its predictions are correct. A well-calibrated model that assigns 70 percent confidence should be correct about 70 percent of the time on comparable cases.
Why it matters: Calibration matters when model confidence affects escalation, human review, risk thresholds, or automated decision-making.
Causal inference
Causal inference is the study of reasoning about cause and effect rather than mere statistical association. It includes formal methods for interventions, counterfactuals, confounding, and causal graphs.
Why it matters: Causal reasoning matters when the question is what will happen if an action is taken, not simply what variables tend to occur together.
Computer vision
Computer vision is the field of AI concerned with extracting useful information from images and video. Tasks include recognition, detection, segmentation, tracking, pose estimation, depth estimation, and visual reasoning.
Why it matters: Computer vision powers applications from industrial inspection and medical imaging to robotics, autonomous vehicles, remote sensing, and creative tools.
Computer vision foundation model
A computer-vision foundation model is trained on broad visual data and intended to transfer across many downstream vision tasks through prompting, fine-tuning, adapters, or task heads.
Why it matters: These models bring the foundation-model paradigm beyond language into segmentation, detection, recognition, multimodal reasoning, and visual generation.
Connectionism
Connectionism is an approach to intelligence based on networks of simple computational units whose collective behavior learns representations and mappings from data.
Why it matters: Modern neural networks and deep learning descend from connectionist ideas, in contrast with traditions that emphasize explicit symbols and hand-authored rules.
Constraint satisfaction problem
A constraint satisfaction problem, or CSP, asks for assignments to variables that satisfy a set of constraints. Scheduling, configuration, timetabling, and logic puzzles are common examples.
Why it matters: CSP methods are a reminder that important AI problems can be solved through structured search and reasoning rather than statistical learning.
Continual learning
Continual learning studies systems that learn from a stream of tasks or data over time while retaining useful knowledge from earlier experience.
Why it matters: A central challenge is catastrophic forgetting, where adaptation to new information degrades performance on earlier tasks.
Convolutional neural network
A convolutional neural network, or CNN, is a neural architecture that applies learned filters across local regions of data, traditionally images, to build increasingly abstract representations.
Why it matters: CNNs drove major advances in image recognition and remain important even as Transformer-based vision architectures have grown more common.
Cynthia Breazeal
Cynthia Breazeal is a roboticist known for pioneering social robotics and for the Kismet project at MIT, which explored how robots could use expressive behavior and social interaction as part of intelligent behavior.
Why it matters: Her work helped establish human-robot interaction and social intelligence as central research problems in robotics.
DARPA Grand Challenge
The DARPA Grand Challenge was a series of autonomous-vehicle competitions intended to accelerate robotic ground-vehicle technology. No team completed the 2004 desert course; in 2005, Stanford's vehicle Stanley won a 132-mile challenge.
Why it matters: The competition helped catalyze a research and engineering community around autonomous driving and demonstrated the value of ambitious public benchmarks for integrated AI systems.
Dartmouth Summer Research Project on Artificial Intelligence
The Dartmouth Summer Research Project on Artificial Intelligence was the 1956 workshop organized around a proposal by John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon. The proposal introduced the term artificial intelligence and helped establish the field as a distinct research agenda.
Why it matters: The workshop is a key historical reference point because AI was framed broadly from the beginning around learning, language, abstraction, reasoning, creativity, and machine intelligence.
Data augmentation
Data augmentation expands a training set by creating transformed or synthetic variants of examples while attempting to preserve the task-relevant label or meaning.
Why it matters: Augmentation can improve generalization, robustness, and sample efficiency, but poor transformations can also teach the model unrealistic invariances.
Data drift
Data drift is change over time in the statistical properties of inputs, labels, behaviors, sensors, or environments that a deployed model encounters.
Why it matters: Monitoring drift helps identify when an AI system may no longer be operating under the assumptions used during development and validation.
David Silver
David Silver is a reinforcement-learning researcher known for leading major work on AlphaGo and subsequent self-play systems at DeepMind.
Why it matters: His work helped demonstrate how deep learning, search, reinforcement learning, and self-play could be combined to solve difficult sequential decision problems.
Decision tree
A decision tree predicts an outcome by recursively splitting examples according to feature values. Trees can perform classification or regression and can be combined into ensembles.
Why it matters: Decision trees remain important because they are flexible, relatively interpretable, and foundational to methods such as random forests and gradient-boosted trees.
Deep Blue
Deep Blue was IBM's chess system that defeated reigning world champion Garry Kasparov in a six-game match in 1997 under standard tournament controls.
Why it matters: Deep Blue is a landmark in game-playing AI and illustrates the power of search, evaluation functions, specialized hardware, and domain engineering before the deep-learning era.
Deep learning
Deep learning is machine learning based on neural networks with multiple layers of learned representations. It is used across vision, speech, language, control, scientific modeling, and multimodal systems.
Why it matters: Deep learning reduced the need for hand-engineered features in many domains by allowing representations to be learned directly from data.
Deep Q-Network
The Deep Q-Network, or DQN, combined Q-learning with deep neural networks to learn policies directly from high-dimensional visual input on Atari games. The work was published in Nature in 2015.
Why it matters: DQN was a landmark demonstration that deep representation learning and reinforcement learning could be combined successfully in a common architecture.
Demis Hassabis
Demis Hassabis is a computer scientist and neuroscientist who co-founded DeepMind and has been closely associated with projects including DQN, AlphaGo, and AlphaFold.
Why it matters: His work helped push AI toward systems that combine learning, search, scientific application, and large-scale research engineering.
Depth estimation
Depth estimation predicts the distance from a camera to points or surfaces in a scene using one or more images, video, stereo cameras, or other sensors.
Why it matters: Depth is important for robotics, navigation, augmented reality, autonomous driving, and 3D scene understanding.
Diffusion model
A diffusion model is a generative model trained to reverse a process that gradually adds noise to data. Generation begins from noise and iteratively denoises toward a sample.
Why it matters: Diffusion methods became central to high-quality image generation and have also been applied to audio, video, 3D content, molecular design, and other domains.
Digital twin
A digital twin is a digital representation of a physical asset, process, or system that is kept connected to real-world data and can support monitoring, simulation, prediction, or decision support.
Why it matters: Digital twins often combine simulation, sensor data, optimization, and AI, especially in manufacturing, infrastructure, logistics, and complex engineered systems.
Distribution shift
Distribution shift occurs when the data encountered after deployment differs in meaningful ways from the data on which a model was trained or validated.
Why it matters: Many AI failures are not caused by a broken algorithm but by a changed environment, population, sensor, workflow, or operating condition.
ELIZA
ELIZA was a 1960s program created by Joseph Weizenbaum that used pattern matching and scripted transformations to simulate text conversation, most famously through the DOCTOR script.
Why it matters: ELIZA is historically important because users could attribute understanding to a system whose underlying mechanism was comparatively simple, a phenomenon still relevant to human interpretation of conversational AI.
Embodied AI
Embodied AI studies intelligent systems whose learning and decision-making are grounded in interaction with a physical or simulated environment through perception and action.
Why it matters: Embodiment makes intelligence a closed-loop problem: sensing, acting, physical constraints, uncertainty, and consequences all matter.
Ensemble learning
Ensemble learning combines multiple models to produce a final prediction, often reducing variance or improving robustness compared with a single model.
Why it matters: Random forests, boosting, bagging, and model averaging are common examples.
Evolutionary algorithm
An evolutionary algorithm optimizes candidate solutions through mechanisms inspired by biological evolution, such as variation, selection, mutation, and recombination.
Why it matters: Evolutionary methods are useful when gradients are unavailable, objectives are irregular, or solutions must be searched over structures rather than continuous parameters.
Expert system
An expert system is an AI program that applies an explicit knowledge base and inference rules to solve problems in a specialized domain.
Why it matters: Expert systems were a major commercial form of AI before the modern machine-learning era and remain important historically because they made knowledge representation and explanation explicit.
Explainable AI
Explainable AI, or XAI, refers to methods and practices intended to make an AI system's behavior, predictions, or internal factors more understandable to people.
Why it matters: Explanation can support debugging, oversight, compliance, scientific insight, and trust, but an explanation method may itself be incomplete or misleading.
Fairness
AI fairness concerns how automated systems distribute errors, opportunities, resources, burdens, or treatment across people and groups.
Why it matters: Different fairness definitions can conflict, so fairness is not a single metric. It requires decisions about context, measurement, affected populations, and acceptable tradeoffs.
Feature engineering
Feature engineering is the process of designing or transforming input variables so that a learning algorithm can use domain-relevant information more effectively.
Why it matters: Deep learning reduced the amount of manual feature engineering in some domains, but data representation remains a central part of practical machine learning.
Federated learning
Federated learning trains or updates models across multiple devices or organizations while keeping raw data decentralized rather than collecting all examples in one central store.
Why it matters: It can reduce some data-sharing requirements, though it does not by itself guarantee privacy, fairness, security, or reliable learning across heterogeneous participants.
Fei-Fei Li
Fei-Fei Li is a computer-vision researcher closely associated with the creation and development of ImageNet, the large labeled image dataset and challenge that helped accelerate modern visual recognition research.
Why it matters: Her work helped make data scale and benchmark design central to the deep-learning revolution in computer vision.
Foundation model
A foundation model is a model trained on broad data at scale that can be adapted or prompted for many downstream tasks rather than being built for a single narrow application.
Why it matters: Foundation models can exist in language, vision, audio, biology, robotics, and multimodal domains.
Frank Rosenblatt
Frank Rosenblatt was a psychologist and computer scientist who developed the perceptron and published a foundational 1958 account of a learning model inspired by neural organization.
Why it matters: The perceptron became an important ancestor of modern neural networks and a focal point in early debates about machine learning.
Fuzzy logic
Fuzzy logic represents degrees of membership or truth rather than restricting propositions to strictly true or false values. It is commonly associated with Lotfi Zadeh's work on fuzzy sets.
Why it matters: Fuzzy systems have been used in control, decision support, consumer electronics, and industrial systems where concepts have gradual boundaries.
Gender Shades
Gender Shades is a research project led by Joy Buolamwini, with a 2018 paper co-authored with Timnit Gebru, that evaluated intersectional accuracy disparities in commercial gender-classification systems.
Why it matters: The project became a landmark example of algorithmic auditing and showed why aggregate accuracy can hide large subgroup performance differences.
General Problem Solver
General Problem Solver, developed by Allen Newell, Herbert A. Simon, and J. C. Shaw, was an early program that attempted to solve different formal problems using general search and means-ends analysis.
Why it matters: It represents an important early attempt to separate a general problem-solving strategy from the details of a single task.
Generative adversarial network
A generative adversarial network, or GAN, trains a generator and discriminator in opposition: the generator creates samples while the discriminator tries to distinguish generated samples from real ones.
Why it matters: GANs reshaped generative modeling in the 2010s and influenced image synthesis, representation learning, and research on synthetic media.
Generative AI
Generative AI refers to models that produce new content or structured outputs such as text, images, audio, video, code, 3D assets, molecules, or other data.
Why it matters: Generative AI includes large language models but is much broader, encompassing diffusion models, GANs, autoregressive models, multimodal models, and domain-specific scientific systems.
Genetic algorithm
A genetic algorithm is an evolutionary optimization method that represents candidate solutions as encodings that can be selected, mutated, and recombined across generations.
Why it matters: Genetic algorithms are used for search and optimization problems where direct analytic or gradient-based methods may be impractical.
Geoffrey Hinton
Geoffrey Hinton is a computer scientist whose work on neural networks, distributed representations, backpropagation, and deep learning helped establish the modern connectionist approach to AI.
Why it matters: Hinton shared the 2018 ACM A.M. Turing Award with Yoshua Bengio and Yann LeCun for conceptual and engineering breakthroughs in deep neural networks.
Gradient descent
Gradient descent is an optimization method that adjusts model parameters in the direction that reduces a differentiable loss function. Variants such as stochastic gradient descent operate on subsets of training data.
Why it matters: Gradient-based optimization is a central engine of modern neural-network training.
Herbert A. Simon
Herbert A. Simon was a scientist whose work with Allen Newell helped establish symbolic AI, computational theories of problem solving, and early systems such as Logic Theorist and General Problem Solver.
Why it matters: Simon connected AI with cognitive science, decision-making, organizational theory, and the study of bounded rationality.
Heuristic
A heuristic is a rule, estimate, or learned function used to guide search or decision-making toward promising options without exhaustively exploring every possibility.
Why it matters: Heuristics are central to classical AI search and remain important in planning, optimization, games, and hybrid learned systems.
Human in the loop
Human-in-the-loop AI keeps people involved in parts of an automated process such as labeling, review, approval, escalation, correction, or policy-setting.
Why it matters: Human involvement can improve oversight, but simply adding a person does not solve problems if the human lacks time, information, authority, or a realistic ability to detect errors.
Ian Goodfellow
Ian Goodfellow is a machine-learning researcher best known as the lead author of the 2014 paper introducing generative adversarial networks.
Why it matters: GANs became one of the most influential generative-modeling frameworks of the deep-learning era.
Image classification
Image classification assigns one or more labels to an image based on its visual content.
Why it matters: Classification was a major benchmark task in the rise of deep computer vision, especially through ImageNet.
Image segmentation
Image segmentation divides an image into meaningful regions or assigns labels to pixels so that objects, surfaces, tissues, or other structures can be separated from one another.
Why it matters: Segmentation is a foundational vision task in medical imaging, autonomous systems, mapping, editing, robotics, and scientific analysis.
ImageNet
ImageNet is a large image dataset organized around WordNet categories and associated with the ImageNet Large Scale Visual Recognition Challenge. It became a central benchmark for visual recognition research.
Why it matters: ImageNet demonstrated the importance of large labeled datasets and helped create the conditions for breakthroughs such as AlexNet.
ImageNet Large Scale Visual Recognition Challenge
The ImageNet Large Scale Visual Recognition Challenge, or ILSVRC, was a competition built around large-scale visual recognition tasks using the ImageNet dataset.
Why it matters: The challenge created a shared measurement regime that made rapid progress visible and helped establish deep learning as the dominant approach in computer vision after 2012.
Imitation learning
Imitation learning trains an agent to reproduce behavior demonstrated by an expert rather than learning only from a hand-designed reward signal.
Why it matters: It is useful in robotics, autonomous driving, and other domains where expert trajectories are available but reward design is difficult.
Instance segmentation
Instance segmentation identifies the pixels belonging to each individual object instance in an image, distinguishing separate objects even when they share the same class.
Why it matters: It combines object detection and segmentation and is useful in robotics, autonomous systems, inspection, medical imaging, and scene understanding.
Interpretability
Interpretability is the degree to which people can understand relevant aspects of how a model represents information, reaches a result, or behaves under different conditions.
Why it matters: Interpretability can refer to inherently understandable models or to techniques that probe complex systems after training.
Inverse reinforcement learning
Inverse reinforcement learning attempts to infer the objective or reward function that could explain demonstrated behavior.
Why it matters: It is useful when desirable behavior is easier to demonstrate than to specify numerically.
John McCarthy
John McCarthy was a computer scientist who coined the term artificial intelligence in the proposal for the 1956 Dartmouth Summer Research Project on Artificial Intelligence and made major contributions to symbolic AI and programming languages.
Why it matters: McCarthy helped define AI as a distinct research field and shaped its early agenda around reasoning, representation, and machine intelligence.
Joy Buolamwini
Joy Buolamwini is a computer scientist and researcher whose Gender Shades work evaluated demographic performance disparities in commercial facial-analysis systems.
Why it matters: Her work helped move algorithmic auditing, dataset representation, and intersectional evaluation into the center of AI accountability discussions.
Judea Pearl
Judea Pearl is a computer scientist whose work on probabilistic reasoning, Bayesian networks, and causal inference transformed how AI systems represent uncertainty and reason about cause and effect.
Why it matters: Pearl received the 2011 ACM A.M. Turing Award for fundamental contributions to artificial intelligence through probabilistic and causal reasoning.
Kismet
Kismet was a social robot developed at MIT by Cynthia Breazeal and colleagues to study expressive, socially responsive interaction between people and robots.
Why it matters: Kismet helped establish social robotics as a field concerned not only with task execution but also with communication, affective cues, turn-taking, and human expectations.
Knowledge graph
A knowledge graph represents entities and relationships as a graph so that information can be queried, connected, and reasoned over.
Why it matters: Knowledge graphs support search, recommendation, data integration, semantic applications, and hybrid systems that combine symbolic structures with machine learning.
Knowledge representation
Knowledge representation studies how facts, concepts, relationships, rules, events, and uncertainty can be encoded so that machines can reason with them.
Why it matters: It is one of the oldest core problems in AI and remains relevant to knowledge graphs, ontologies, planning, reasoning, and hybrid neuro-symbolic systems.
Logic Theorist
Logic Theorist, created in the 1950s by Allen Newell, Herbert A. Simon, and J. C. Shaw, was an early AI program that proved theorems from Principia Mathematica using heuristic search.
Why it matters: It is one of the earliest demonstrations of symbolic problem solving and heuristic search in AI.
LSTM
Long short-term memory, or LSTM, is a recurrent neural-network architecture designed to preserve useful information over longer sequences through gated memory mechanisms.
Why it matters: LSTMs became widely used for speech, handwriting, language, and time-series tasks before Transformers became dominant in many sequence domains.
Machine learning
Machine learning is the study and use of algorithms that improve performance on tasks by learning patterns from data or experience rather than relying only on explicit hand-written rules.
Why it matters: Machine learning includes supervised, unsupervised, self-supervised, reinforcement, online, active, federated, and many other learning paradigms.
Markov decision process
A Markov decision process, or MDP, formalizes sequential decision-making using states, actions, transition probabilities, rewards, and a policy.
Why it matters: MDPs provide the mathematical foundation for much of reinforcement learning and planning under uncertainty.
Marvin Minsky
Marvin Minsky was a pioneering AI researcher and co-founder of the MIT Artificial Intelligence Laboratory. His work ranged across symbolic reasoning, neural networks, perception, robotics, and theories of mind.
Why it matters: Minsky helped shape early AI as an intellectual field and influenced decades of research on representation, reasoning, and cognition.
Marvin Minsky and Seymour Papert's Perceptrons
Perceptrons was a 1969 book by Marvin Minsky and Seymour Papert that analyzed the capabilities and limitations of single-layer perceptron systems.
Why it matters: The book became part of the historical debate over neural-network research and is often discussed in accounts of the field's shifts between symbolic and connectionist approaches.
Minimax
Minimax is a decision rule for two-player zero-sum games in which one player selects moves to maximize value while assuming the opponent selects moves to minimize it.
Why it matters: It is a foundational game-playing method and illustrates adversarial search before learning-based game agents.
Model predictive control
Model predictive control, or MPC, repeatedly optimizes a sequence of future control actions using a model of system dynamics, executes part of the plan, then replans as new state information arrives.
Why it matters: MPC is a major bridge between control theory and autonomous systems and is often combined with learned models or perception systems.
Model-based reinforcement learning
Model-based reinforcement learning uses or learns a model of how the environment changes so the agent can plan by predicting consequences before acting.
Why it matters: Model-based approaches can improve sample efficiency and support planning, but performance depends on the accuracy and usefulness of the learned model.
Model-free reinforcement learning
Model-free reinforcement learning learns a policy or value function without requiring an explicit predictive model of environment dynamics.
Why it matters: Many influential deep-RL methods are model-free, though they can require large amounts of interaction data.
Monte Carlo tree search
Monte Carlo tree search, or MCTS, builds a search tree selectively by using repeated simulations to estimate the value of actions and focus computation on promising branches.
Why it matters: MCTS became especially prominent through game-playing systems such as AlphaGo, where search was guided by learned policy and value networks.
Multilayer perceptron
A multilayer perceptron, or MLP, is a feed-forward neural network composed of layers of learned weighted transformations and nonlinear activation functions.
Why it matters: MLPs are a basic building block of deep learning and also appear inside more complex architectures such as Transformers.
Multimodal AI
Multimodal AI learns from or operates across more than one type of data, such as text, images, audio, video, sensor signals, actions, or structured data.
Why it matters: Multimodal systems can connect perception, language, and action and are important for robotics, assistants, scientific systems, and human-computer interaction.
MYCIN
MYCIN was a 1970s expert system developed at Stanford to reason about bacterial infections and antibiotic treatment using an explicit rule base and certainty factors.
Why it matters: MYCIN became a landmark example of expert-system design, knowledge engineering, and the challenge of deploying AI in high-stakes domains.
Natural language processing
Natural language processing, or NLP, is the field concerned with computational analysis, understanding, generation, and transformation of human language.
Why it matters: NLP includes parsing, information extraction, translation, search, question answering, dialogue, text classification, language modeling, and speech-language interfaces.
Neural network
A neural network is a parameterized computational model built from layers or networks of simple units that transform inputs into outputs. Learning adjusts the parameters to improve performance on data or experience.
Why it matters: Neural networks are used in perception, language, control, generative modeling, scientific computing, and many other AI domains.
Neuro-symbolic AI
Neuro-symbolic AI combines learned neural representations with explicit symbolic structures, logic, search, constraints, or knowledge representations.
Why it matters: The goal is often to combine the flexibility and perception capabilities of neural models with the compositional structure and explicit reasoning of symbolic systems.
Object detection
Object detection identifies and localizes instances of objects in images or video, commonly by predicting classes and bounding boxes.
Why it matters: Detection is a core perception task for autonomous vehicles, robotics, surveillance, retail analytics, industrial inspection, and many other applications.
OCR
Optical character recognition, or OCR, converts text in images or scanned documents into machine-readable characters and structure.
Why it matters: Modern OCR often combines image processing, computer vision, sequence modeling, language models, and document-layout understanding.
Ontology
An ontology is a formal representation of concepts, categories, properties, and relationships within a domain.
Why it matters: Ontologies support semantic interoperability, knowledge graphs, reasoning, data integration, and systems that need explicit shared meaning.
Optical flow
Optical flow estimates apparent motion of pixels or visual features between frames in a sequence.
Why it matters: It supports tracking, video analysis, robotics, autonomous navigation, motion estimation, and scene understanding.
Out-of-distribution generalization
Out-of-distribution generalization is the ability of a model to perform well on examples or environments that differ meaningfully from the training distribution.
Why it matters: It is a central deployment challenge because real-world systems inevitably encounter novel populations, conditions, sensors, tasks, and edge cases.
Overfitting
Overfitting occurs when a model learns training-specific patterns or noise so strongly that performance does not generalize well to new data.
Why it matters: Preventing overfitting motivates validation, regularization, data augmentation, early stopping, appropriate model capacity, and careful evaluation.
Perceptron
A perceptron is an early learning model that computes a weighted combination of inputs and applies a decision rule. Frank Rosenblatt's work in the 1950s made the perceptron a foundational model for machine learning and neural networks.
Why it matters: The perceptron is historically important both as a practical learning algorithm and as a precursor to multilayer neural networks.
Planning
Planning is the problem of finding a sequence or policy of actions that transforms an initial state toward a goal while respecting constraints and predicted consequences.
Why it matters: Planning is central to robotics, logistics, games, operations research, autonomous agents, and classical AI.
Policy
A policy is the rule or learned mapping an agent uses to select actions from states, observations, or histories.
Why it matters: Learning a good policy is the central objective of many reinforcement-learning and control problems.
Pose estimation
Pose estimation predicts the location or configuration of structured keypoints, such as human joints, hands, faces, or articulated objects, from images or video.
Why it matters: It supports human-computer interaction, sports analysis, robotics, animation, ergonomics, and safety applications.
Probabilistic graphical model
A probabilistic graphical model represents a probability distribution using a graph whose structure captures conditional dependencies among variables.
Why it matters: Bayesian networks and Markov random fields are major examples and provide a bridge between probability, structure, and inference.
Q-learning
Q-learning is an off-policy reinforcement-learning algorithm that learns the expected long-term return of taking an action in a state and then acting optimally afterward.
Why it matters: It is a foundational value-based method and the conceptual basis of systems such as Deep Q-Networks.
Random forest
A random forest is an ensemble of decision trees trained on randomized subsets of examples and features, with predictions combined across trees.
Why it matters: Random forests remain strong baselines for structured tabular data and are often robust with limited feature preprocessing.
Recurrent neural network
A recurrent neural network, or RNN, processes sequences by carrying forward a hidden state that depends on previous inputs.
Why it matters: RNNs were central to speech, language, handwriting, and time-series modeling before attention-based architectures displaced them in many applications.
Regularization
Regularization is any technique that discourages a model from fitting training data in ways that reduce generalization, often by constraining parameters, adding noise, limiting capacity, or stopping training early.
Why it matters: Regularization helps balance model flexibility against the risk of overfitting.
Reinforcement learning
Reinforcement learning is a computational approach in which an agent learns how to act by interacting with an environment and receiving rewards or other evaluative feedback.
Why it matters: Unlike ordinary supervised learning, reinforcement learning focuses on sequential decisions, delayed consequences, exploration, and policies.
Representation learning
Representation learning allows a model to discover useful features or latent structure from data instead of relying entirely on features designed by people.
Why it matters: Deep learning is powerful in part because layered representations can be learned jointly with the task.
ResNet
ResNet, or residual network, is a deep neural architecture introduced by Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun that uses residual connections to make very deep networks easier to optimize.
Why it matters: Residual connections became a standard design pattern across deep learning, far beyond the original computer-vision setting.
Reward function
A reward function assigns evaluative feedback to states, actions, or outcomes so that a reinforcement-learning agent has a signal about what behavior should be preferred.
Why it matters: Poor reward design can produce unintended strategies, reward hacking, or behavior that optimizes the metric without satisfying the real objective.
Reward hacking
Reward hacking occurs when an agent finds a way to obtain high reward without accomplishing the intended objective, often by exploiting gaps in the reward function or environment.
Why it matters: It illustrates why specifying what a system should optimize can be as difficult as building the optimizer itself.
Richard Sutton
Richard S. Sutton is a computer scientist whose work helped establish reinforcement learning as a distinct modern field, including temporal-difference learning and the textbook Reinforcement Learning: An Introduction with Andrew Barto.
Why it matters: Sutton's work has shaped how researchers think about agents learning predictions, values, and behavior from interaction.
Robotics
Robotics combines sensing, computation, planning, control, mechanics, and interaction to build machines that act in the physical world.
Why it matters: AI contributes perception, decision-making, learning, planning, adaptation, and human-robot interaction, but a capable robot also depends on control, hardware, safety, and systems engineering.
Robustness
Robustness is the ability of an AI system to continue performing acceptably when inputs, conditions, users, environments, or perturbations differ from the clean cases seen during development.
Why it matters: Robustness is broader than average accuracy and is central to deployment in open, changing, adversarial, or safety-critical environments.
Rodney Brooks
Rodney Brooks is a roboticist known for behavior-based robotics and the subsumption architecture, an approach that emphasized layered reactive behaviors rather than relying entirely on centralized symbolic world models.
Why it matters: His work helped shift robotics toward embodied interaction and challenged assumptions that intelligent behavior required complete internal representations before action.
Segment Anything
Segment Anything is a Meta AI project introduced in 2023 that combined a promptable segmentation model with a large segmentation dataset containing more than a billion masks.
Why it matters: The project helped bring foundation-model ideas to computer vision by treating segmentation as a broadly promptable task rather than a separately trained model for each domain.
Self-play
Self-play trains an agent by having it generate experience through competition or interaction with copies or variants of itself.
Why it matters: Self-play can create a curriculum automatically and was central to systems such as AlphaGo Zero and AlphaZero.
Self-supervised learning
Self-supervised learning creates training signals from the structure of unlabeled data itself, such as predicting masked content, future content, transformations, or relationships among views.
Why it matters: It enabled models to learn broadly useful representations from massive datasets without requiring human labels for every example.
Semantic segmentation
Semantic segmentation assigns a class label to each pixel in an image, such as road, sky, person, tumor, or building, without necessarily separating individual objects of the same class.
Why it matters: It is important in autonomous driving, remote sensing, medicine, robotics, industrial inspection, and scene understanding.
Semi-supervised learning
Semi-supervised learning combines a smaller set of labeled examples with a larger pool of unlabeled data.
Why it matters: It is useful when data are plentiful but accurate labels require expensive expert work.
Sensor fusion
Sensor fusion combines information from multiple sensors, such as cameras, radar, lidar, inertial units, GPS, microphones, or proprioceptive sensors, to estimate the state of the world more reliably.
Why it matters: Fusion is central to autonomous vehicles, robotics, navigation, tracking, and systems that must operate when individual sensors are noisy or incomplete.
SHRDLU
SHRDLU was Terry Winograd's early natural-language system that could understand commands and questions about a simplified blocks world, combining parsing, reasoning, and a constrained model of the environment.
Why it matters: SHRDLU became a classic demonstration of how language understanding can look powerful inside a tightly bounded world while struggling to generalize beyond it.
Sim-to-real
Sim-to-real refers to techniques for training or testing an AI system in simulation and transferring the resulting behavior to the physical world.
Why it matters: The core challenge is the reality gap: simulated sensors, physics, environments, and rare events never perfectly match deployment.
SLAM
Simultaneous localization and mapping, or SLAM, is the problem of building or updating a map of an environment while estimating the agent's own position within it.
Why it matters: SLAM is foundational to mobile robotics, drones, autonomous navigation, augmented reality, and systems that operate without a complete prior map.
Speech recognition
Automatic speech recognition, or ASR, converts spoken audio into text or other symbolic representations.
Why it matters: Speech recognition combines signal processing, acoustic modeling, sequence modeling, language modeling, and increasingly end-to-end deep learning.
Stanley
Stanley was Stanford Racing Team's autonomous vehicle that won the 2005 DARPA Grand Challenge, completing a 132-mile desert course.
Why it matters: The project became a landmark in autonomous driving by integrating perception, mapping, planning, vehicle control, and robust systems engineering.
Supervised learning
Supervised learning trains a model from examples paired with target labels or outputs, such as images with categories, transactions with outcomes, or sensor readings with known states.
Why it matters: It remains one of the most widely used learning paradigms when reliable labeled data are available.
Swarm intelligence
Swarm intelligence studies decentralized problem-solving methods inspired by collective behavior in systems such as ant colonies, bird flocks, or particle populations.
Why it matters: Algorithms such as ant-colony optimization and particle-swarm optimization search through coordinated local interactions rather than centralized reasoning.
Symbolic AI
Symbolic AI represents knowledge using explicit symbols, rules, logical expressions, search procedures, or structured relationships that can be manipulated by algorithms.
Why it matters: Symbolic methods dominated much early AI and remain valuable for reasoning, planning, constraint solving, knowledge representation, and hybrid systems.
Timnit Gebru
Timnit Gebru is a computer scientist known for research on algorithmic bias, dataset documentation, computer vision, and the social consequences of large-scale AI systems. She co-authored the Gender Shades study with Joy Buolamwini.
Why it matters: Her work helped establish dataset composition, documentation, and sociotechnical evaluation as central questions in responsible AI.
Training data
Training data are the examples, observations, interactions, labels, or generated experiences used to adjust a model's parameters or policy.
Why it matters: Data composition strongly influences what a model can learn, which populations or conditions it represents well, and what errors it is likely to make.
Transfer learning
Transfer learning reuses representations or parameters learned on one task or dataset as the starting point for another task.
Why it matters: It can reduce data and compute requirements and is foundational to modern pretrained models in vision, language, audio, and multimodal AI.
Transformer
A Transformer is a neural architecture built around attention mechanisms rather than recurrence as the primary way to mix information across a sequence or set.
Why it matters: Introduced in the 2017 paper Attention Is All You Need, Transformers became central to modern language models and also spread into computer vision, audio, robotics, biology, and multimodal systems.
Uncertainty estimation
Uncertainty estimation attempts to quantify how uncertain a model is about a prediction, latent state, or future outcome.
Why it matters: Reliable uncertainty can support abstention, human escalation, risk-sensitive planning, active learning, and safer decision-making.
Underfitting
Underfitting occurs when a model is too limited, poorly trained, or insufficiently informed to capture important patterns even in the training data.
Why it matters: It is the opposite side of the model-complexity problem from overfitting and often appears as weak performance on both training and test data.
Unsupervised learning
Unsupervised learning seeks useful structure in data without target labels supplied for every example, such as clusters, latent factors, densities, or representations.
Why it matters: Many modern representation and generative methods blur the boundary between unsupervised and self-supervised learning.
Value function
A value function estimates expected future return from a state, or from a state-action pair, under a policy or optimal behavior.
Why it matters: Value estimation allows agents to compare consequences that unfold over time and is central to many reinforcement-learning algorithms.
Variational autoencoder
A variational autoencoder, or VAE, is a generative latent-variable model that learns a probabilistic mapping between data and a structured latent space.
Why it matters: VAEs are important for generative modeling, representation learning, anomaly detection, and as components inside larger generative systems.
Visual tracking
Visual tracking estimates the location, identity, or trajectory of objects across successive video frames.
Why it matters: Tracking links individual visual observations over time and is central to surveillance, sports analytics, robotics, autonomous driving, and human-computer interaction.
World model
A world model is an internal predictive model of how an environment behaves, including how states evolve and how actions affect future observations or outcomes.
Why it matters: World models allow agents to simulate possibilities, plan ahead, learn from imagined trajectories, and reason beyond immediate perception.
Yann LeCun
Yann LeCun is a computer scientist whose work on convolutional neural networks and gradient-based learning helped establish modern computer vision and deep learning.
Why it matters: LeCun shared the 2018 ACM A.M. Turing Award with Geoffrey Hinton and Yoshua Bengio for breakthroughs that made deep neural networks a major computing technology.
YOLO
YOLO, short for You Only Look Once, is a family of real-time object-detection approaches originating with a 2016 paper by Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi that framed detection as a single end-to-end prediction problem.
Why it matters: YOLO helped make fast neural object detection practical and influential in robotics, video, inspection, and edge vision.
Yoshua Bengio
Yoshua Bengio is a computer scientist whose research on neural networks, representation learning, and deep learning helped establish the modern learning-based approach to AI.
Why it matters: Bengio shared the 2018 ACM A.M. Turing Award with Geoffrey Hinton and Yann LeCun for breakthroughs in deep neural networks.
The vocabulary matters because different AI systems fail in different ways.
TDA helps organizations understand the software, dependencies, AI-assisted development patterns, ownership, maintainability, and continuity behind systems they increasingly depend on.