AI glossary

A method that reconstructs a three-dimensional model of an object or scene from photographs, videos, scans, or sensor data.

Comparing two or more variations of a product, tooltip, interface, or model to choose the best based on data rather than feel.

The principle that clear people, teams, or organizations should be held accountable for the decisions, errors, and consequences of an AI system.

A quality metric that shows what percentage of predictions the model made correctly.

An approach to speech recognition that associates an audio signal with likely sounds, syllables, or words.

Mathematical functions in a neural network that help it find complex nonlinear dependencies.

A machine learning method in which the model itself selects the most useful examples for human labeling.

A class of reinforcement learning methods where one component selects actions and another evaluates their utility.

The physical components of a robot or system that translate an algorithm's command into an action: moving, pressing, turning, or grasping.

An optimization algorithm that automatically adjusts the neural network training step and reduces dependence on manual selection of the learning rate.

An adaptive optimization algorithm that adjusts the learning step separately for each model parameter.

One of the most popular neural network training algorithms that adaptively adjusts parameter updates.

A statistical metric that evaluates the quality of a regression model, adjusted for the number of features.

Ways to fool a model with specially selected input data that looks normal to a human but confuses the algorithm.

An approach in robotics in which the system learns to understand what actions are possible with objects and the environment.

A modeling method where the behavior of a system is described through many individual agents with their own rules.

AI systems that can plan actions, use tools, and complete a task in a few steps with limited human intervention.

Sets of libraries and rules that help create AI agents with tools, memory, roles and execution scripts.

A data grouping method that gradually groups similar objects into clusters.

Specialized computing devices that speed up the training and execution of artificial intelligence models.

A European regulation that introduces risk-based rules for the development, sale and use of AI systems.

AI-based programs that can perform tasks through planning, memory, tools, and sequencing.

A field of study and practice that attempts to make AI behavior safe, beneficial, and consistent with human intentions.

A situation where an AI system systematically produces more favorable or unfavorable results for specific groups, cases or traits.

A direction where AI is used for environmental, social and economic problems, and also evaluates its own impact on resources.

Software interfaces through which applications connect to AI models, services for generating, recognizing or analyzing data.

Creation of pictures, illustrations, concepts and visual materials using neural networks based on text, reference or settings.

AI-powered digital assistants that answer questions, help with tasks, texts, searches, code or work organization.

Reviewing an AI system for quality, risk, safety, fairness, compliance, and suitability for real-world use.

The process of regularly reviewing an AI system, its data, behavior, documentation, and consequences of use.

US policy framework on protecting people when using automated systems and artificial intelligence.

A set of measures that limit the capabilities of an AI system so that it does not exceed safe behavior limits.

Expanding access to AI technologies for people, small businesses, developers, education and organizations without large resources.

Panels of experts or internal committees that evaluate risks, rules, and controversial decisions in the development and implementation of AI.

Sets of principles and rules that help develop and use AI safely, fairly and transparently.

A system for storing, reusing and managing data features that are used in machine learning.

Application of artificial intelligence to analyze, forecast and reduce climate risks.

Texts, images, audio, video, code or other materials created in whole or in part using artificial intelligence.

Structures, policies and processes that help an organization securely develop, implement and control AI.

Rules, filters and technical barriers that help keep the AI ​​system within safe and acceptable limits.

Errors in which the AI ​​confidently produces fictitious, inaccurate or unconfirmed information.

Principles for the safe, fair and responsible use of AI in medicine and related services.

The use of artificial intelligence in defense, intelligence, cyber and autonomous military systems.

Laws and regulations that govern the creation, sale, implementation and use of artificial intelligence systems.

Ability to understand the capabilities, limitations, risks and rules for the safe use of artificial intelligence.

Platforms where you can find, compare, buy or connect AI tools, models, agents and ready-made solutions.

Model control processes throughout the entire life cycle: from selection and training to launch, monitoring, updating and decommissioning.

A platform or directory where artificial intelligence models are published, searched, downloaded, tested or connected.

A system for recording models, their versions, statuses, metrics, owners and change history.

Numerical parameters of a neural network that determine how the model transforms input data into a response.

Creating melodies, arrangements, beats, background music or vocal fragments using neural networks.

Coordinating models, tools, data, and process steps so that the AI ​​system performs a complex task in a controlled manner.

A set of rules, goals, and decisions by an organization or government about how to develop, use, and control AI.

Government, industry, and corporate regulations that limit and guide the use of artificial intelligence.

An area that deals with reducing the risks from errors, abuse, unpredictable behavior, and harmful effects of AI systems.

Protecting AI systems from attacks, leaks, abuse, unauthorized access and dangerous behavior.

A hypothetical moment when the development of artificial intelligence becomes so fast and powerful that the consequences are difficult to predict.

A document in which an organization discloses information about the operation, risks, data, limitations and controls of AI systems.

Tags or signals that help determine that content was created or modified by artificial intelligence.

A period of decline in interest, funding and expectations for artificial intelligence after inflated promises and weak results.

Using AI to perform repetitive tasks, transfer data between services, and support multi-step work scenarios.

An access model in which artificial intelligence functions are provided through a cloud service, API, or ready-made web interface.

British mathematician and one of the key scientists whose ideas became the basis of computer science and thinking about artificial intelligence.

An approach in which the decisions of algorithms must be controllable, verifiable, and associated with responsible people or organizations.

Techniques that help reduce unfair or systematic biases in data, models, and automated decisions.

A situation where an automated system unfairly makes people worse off because of attributes, data, or decision rules.

The ability to explain how an algorithm works, what data it uses, and why it makes certain decisions.

Scientist, teacher and entrepreneur known for his contributions to machine learning, online education and the popularization of AI.

Methods of searching for unusual events, objects, or patterns that differ markedly from normal data behavior.

Search for unusual actions, events, or network activity that may indicate an attack, leak, or compromise.

An artificial intelligence company known for its Claude family of models and its focus on AI security, controllability, and reliability.

A data streaming platform that helps systems share events in real time.

Design and creation of software interfaces through which services exchange data and functions.

Addresses or API routes through which the application accesses a specific service function.

A method of finding frequently occurring relationships and rules in data sets, especially shopping carts and transactions.

A hypothetical AI that can solve a wide range of tasks at or above human level, rather than just a narrow, specialized task.

The general name for technologies that enable computers to perform tasks typically requiring human intelligence.

Machine learning models inspired by a network of interconnected nodes that learn to transform data into the desired output.

A hypothetical AI that is significantly superior to humans in almost all intellectual tasks.

A specialized chip designed to perform a specific type of calculation as efficiently as possible.

Machine learning methods that look for consistent relationships between elements, events, or features in data.

A reinforcement learning method where multiple agents learn and update a common model in parallel.

Neural network components that help the model identify important parts of the input data and take into account the relationships between them.

A classifier evaluation metric that shows how well the model separates positive and negative examples at different thresholds.

A task in which the AI ​​determines the type of sound, event, speech, music, noise, or other audio condition.

Methods for analyzing, enhancing, transforming and extracting information from audio data.

A technology that overlays digital objects, cues, or data onto an image of the real world.

Verifying the identity of a user, service or device before accessing the system.

Checking what data and actions are allowed to an already authenticated user or service.

Neural network models that learn to compress data into a compact representation and restore it back.

An area of ​​AI where systems attempt to draw logical conclusions, prove statements, and solve problems using formal rules.

Technology that turns spoken speech into text using audio processing and language models.

Transferring repetitive actions to a system that performs them according to rules, data or AI decisions without constant manual intervention.

An approach in which the system helps automatically select models, features, parameters and training stages.

The ability of a robot, drone, car or other system to independently navigate and move in space.

Models that predict the next element in a sequence based on previous elements.

A neural network training algorithm that calculates how to change model weights to reduce error.

A simple way to think of text as a collection of words and their frequencies without regard to order or grammar.

An ensemble learning method where multiple models are trained on different subsamples of data and then their responses are combined.

A method in deep neural networks that stabilizes the distribution of values ​​within a layer and helps speed up learning.

The number of examples the model processes in one training or inference step.

A way to update the probability estimate as new data becomes available.

A graphical model that shows probabilistic relationships between events or features.

A method for finding good parameters when each check is expensive or time consuming.

A transformer-based language model that takes into account the left and right context of a word.

A systematic error that causes a model to perform less well for some users, data, or scenarios.

A balance between a model that is too simple and a model that overtrains on random details in the data.

Very large, fast, or diverse data sets that require special storage and processing approaches.

Tools and architectures for storing, transmitting, processing and analyzing large data streams.

Protect access using biometrics: face, voice, fingerprint, iris or behavior.

A metric that compares machine translation or generated text to reference versions.

Using a distributed ledger to record data, rights, model provenance, or operations of AI systems.

An ensemble learning method, where models are built sequentially and each new one corrects the errors of the previous ones.

A graph that shows the distribution of data through the median, quartile limits, and outliers.

A way to break text into tokens, gradually combining frequently occurring fragments of characters.

A deep learning framework known for early use in computer vision and neural network experiments.

The branch of mathematics about derivatives, integrals and changes in quantities that underlies the training of neural networks.

Matching the model's confidence with the actual probability of the correct answer.

California Consumer Rights Act Regarding Personal Information Collected by Businesses.

Gradual launch of a new version of the system on a small proportion of users before a full rollout.

A neural network architecture that attempts to better account for parts of an object and their relative positions.

Cumulative emissions associated with training, running, and maintaining AI systems.

An approach in which a system solves a new problem by finding similar past cases and adapting their solutions.

The problem is when a model, after training a new task, sharply loses quality on old tasks.

A statistical principle that explains why the average of many independent observations often behaves predictably.

A way to describe intermediate steps in solving a problem to improve the testability and controllability of the answer.

A thought experiment about whether a system's correct answer implies a real understanding of the meaning.

A situation where there are many more examples in the data than others.

A machine learning problem where an object needs to be classified into one or more classes.

Services that provide access to models, compute, storage, APIs, and deployment tools through the cloud.

Using remote servers and services to store, process data and run applications.

A method for finding groups of similar objects without predefined classes.

A set of methods that automatically group similar objects based on characteristics.

AI models that create, augment, explain, or correct program code on demand.

Models of the structure of an intelligent system that describe memory, goals, perception and decision making.

An approach to creating systems that mimic individual elements of human analysis and decision making.

A company and platform developing language models and enterprise AI solutions.

An online environment for running code, experimenting with data, and training models in interactive notebooks.

A recommendation method that relies on similarities between users, products, and their past actions.

Ways to represent colors numerically, such as through brightness, hue, or image channels.

A branch of mathematics about functions of a complex variable that is found in signals, modeling, and some numerical methods.

Compliance of product, data and processes with internal regulations, laws, standards and industry requirements.

A set of standards and guidelines that help manage quality, risk, and the responsible use of technology.

An estimate of how much time, memory, or operations an algorithm requires as the problem size grows.

Modeling the movement of liquids and gases using numerical methods and calculations.

The idea of describing an agent's beliefs, intentions, and behavior through computable models.

Models that balance size, data, and compute budget for the best results at an affordable price.

Changes in the relationship between input data and the correct answer over time.

Representation of knowledge in the form of nodes and connections between concepts.

The probability of an event given that another event has already occurred or is known.

A table that shows which classes the model recognized correctly and where it mixed up the answers.

A debatable topic is whether an artificial system can have subjective experience or only simulate responses.

A class of generative models that strive to produce high-quality results in fewer steps.

A method of training AI where a model's behavior is guided through a set of principles and rules.

A class of problems where you need to find a solution that simultaneously respects a set of rules and conditions.

Packaging the application and its dependencies into an isolated environment that is easier to port and run.

A recommendation method that selects objects based on their characteristics and user interests.

The amount of text, images, or other data that the model can consider in a single request.

An approach in which the model gradually learns from new data without forgetting previous knowledge.

The practice of automatically checking, building, and deploying changes to code, data, or models.

Systems that measure the state of an object and select actions to achieve the desired behavior.

Systems that conduct a dialogue with the user through text or voice.

A neural network architecture well suited for image, video, and spatial features.

The U.S. Children's Online Privacy Protection Act, especially when collecting data from users under 13 years of age.

The task is to determine which words or phrases in the text refer to the same object.

A statistical relationship between two quantities, showing how they change together.

Table of mutual correlations between several characteristics.

An approach where different model errors have different costs and are taken into account during training or evaluation.

An error function often used in classification problems to evaluate probabilistic predictions.

Test the model on multiple data partitions to more robustly assess quality.

An approach where the model is trained from simpler examples to more complex ones.

Collection and analysis of information about cyber threats, attackers, vulnerabilities and signs of attacks.

Laws, standards and industry requirements governing the protection of digital systems and data.

Model and product for generating images from text descriptions.

Adding explanatory labels to data so the model can learn from examples.

Artificial increase in the training sample due to transformations or generation of new data options.

A system for recording data sets, their owners, descriptions, sources, quality and access rights.

Physical infrastructure with servers, networks, cooling and power to run digital services.

Removing, correcting, or normalizing errors, omissions, duplicates, and strange values in data.

Changing the distribution of input data after running the model.

Rules, roles and processes that define how data is collected, stored, used and controlled.

Receiving data from sources and transferring it to a warehouse, pipeline, or processing system.

Combining data from different sources into a coherent structure for analysis or model training.

Assign labels, classes, or annotations to data to train and validate models.

A storage facility where various data can be stored in raw or poorly processed form for subsequent analysis.

An architectural approach to storing large volumes of heterogeneous data for analytics and AI.

An error occurs when training or testing a model includes information that is not available in real use.

Track the path of data from its source through transformations to a report, model, or solution.

Find patterns, groups, connections and useful signals in large data sets.

A sequence of steps that collects, cleanses, transforms, and feeds data into an analytics or model.

An attack or bug in which harmful data gets into the training and degrades the behavior of the model.

Data preparation before analysis or training: cleaning, transformation, coding and normalization.

Protect personal and sensitive information from unnecessary collection, disclosure and unauthorized use.

Technical and organizational measures that prevent the loss, leakage, misuse or illegal processing of data.

Legal regulations governing the collection, storage, processing, transfer and deletion of personal data.

The degree to which data is complete, accurate, current, consistent, and appropriate for the task.

Data protection from unauthorized access, modification, destruction and leakage.

A requirement that data be subject to the rules of the country or region in which it is located or to which it relates.

Changing the format, structure, or values of data so that it is suitable for analysis, storage, or a model.

Accounting for versions of data sets to understand what data the model was trained or tested on.

The practice of turning data into charts, dashboards, maps and other visual forms so patterns are easier to see.

Methods for choosing charts and visual encodings that make data comparisons clearer and less misleading.

A centralized approach to storing cleaned, consistent data for reporting, analytics and AI workflows.

Database systems and retrieval patterns designed to give AI applications reliable access to useful information.

A clustering quality metric that compares how compact clusters are and how far apart they are from each other.

A density-based clustering algorithm that finds groups of nearby points while marking outliers as noise.

A way to show how a model separates feature space into different predicted classes.

Models that make predictions through a sequence of human-readable questions and branches.

Layered probabilistic neural networks built from stacked representation-learning components.

A family of machine learning methods that use multi-layer neural networks to learn patterns from data.

Reinforcement learning models that use neural networks to estimate the value of actions.

Reinforcement learning that uses deep neural networks to learn decisions from interaction and reward.

Synthetic media that uses AI to imitate a person, voice, face or event.

Google's AI research lab known for work on advanced models, reinforcement learning and scientific AI systems.

An AI researcher and entrepreneur associated with the founding and leadership of Google DeepMind.

An NLP task that identifies grammatical relationships between words in a sentence.

A computer vision task that estimates how far objects or surfaces are from the camera.

Summary measures that describe the main properties of a dataset.

Software systems that manage conversations between people and machines.

Equations that describe how quantities change and are often used to model dynamic systems.

An evolutionary optimization method that searches for better solutions by combining candidate variants.

A privacy technique that limits what can be inferred about any one person from aggregate data.

Generative models that create data by learning to reverse a gradual noise-adding process.

The gap between people and organizations that can access digital technology and those that cannot.

Information that helps trace the origin, history and edits of digital content.

Virtual representations of real systems, assets or processes that can be monitored and simulated.

Methods that compress many features into fewer dimensions while preserving useful structure.

A preference-tuning method that trains models directly from chosen and rejected responses.

The study of how meaning is built across sentences, conversations and larger text contexts.

A simulation method that models systems as sequences of events over time.

The branch of mathematics focused on countable structures such as graphs, logic and combinatorics.

False or misleading information spread with the intent to deceive or manipulate.

The use of multiple connected machines to run computation that would be too large for one system.

Storage systems that spread files across multiple machines while presenting them as one system.

Training AI models across several processors, servers or accelerators at the same time.

A container platform that packages applications with their dependencies so they run consistently across environments.

Uncrewed aerial systems that can use AI for navigation, inspection, mapping and perception.

A regularization technique that randomly disables parts of a neural network during training.

The use of dropout to reduce overfitting and make neural networks more robust.

Systems whose state changes over time according to rules or equations.

A training strategy that stops model training when validation quality stops improving.

AI that runs on local devices or nearby hardware instead of relying only on cloud servers.

Computing performed close to where data is produced, reducing latency and data transfer.

A computer vision technique that finds boundaries and sharp changes in images.

Linear algebra concepts that describe important directions and scaling behavior in transformations.

An entrepreneur linked to public AI initiatives, autonomous systems and debates about AI safety and platforms.

The number of dimensions used to represent an item as a vector embedding.

Numeric vector representations that capture meaning, similarity or structure in data.

AI designed to act through a body or agent in a physical or simulated environment.

Capabilities that appear in AI systems at larger scales and were not obvious in smaller versions.

AI methods that attempt to infer emotional states from text, voice, images or behavior.

Machine learning approaches that combine several models to improve accuracy or stability.

A memory concept focused on storing specific events, interactions or experiences.

The design and use of AI systems with attention to fairness, safety, accountability and human impact.

Principles and rules used to guide responsible AI design, deployment and governance.

The review of AI systems for risks such as bias, privacy issues, opacity and harmful outcomes.

The process of choosing AI actions or policies while considering values, harms and responsibilities.

Situations where AI choices involve competing values, risks or stakeholder interests.

Structured approaches for evaluating the social, legal and moral consequences of AI systems.

The field that studies how AI should be designed, governed and used responsibly.

Extract, transform and load workflows that move data from sources into usable analytical systems.

Optimization methods inspired by selection, mutation and recombination.

The possibility that advanced technology could cause irreversible harm to humanity's long-term future.

An iterative statistical method for estimating hidden variables and model parameters.

The concrete EM procedure that alternates between estimating hidden structure and updating parameters.

A metric that measures how well predicted probabilities match actual outcomes.

Rule-based AI systems that encode specialist knowledge to make recommendations or decisions.

The ability to understand why an AI system produced a particular output or decision.

Methods and practices that make AI behavior easier for people to inspect and understand.

The tradeoff between trying new actions and using actions that already seem to work.

The first stage of data analysis focused on understanding patterns, quality issues and hypotheses.

Legal restrictions that limit the transfer of sensitive technologies, models, chips or know-how.

A classification metric that balances precision and recall in one score.

Computer vision technology that identifies or verifies people from facial features.

The goal of reducing unjust or harmful differences in how AI systems treat people or groups.

A deep learning library and educational ecosystem built on top of PyTorch.

The process of creating useful input variables from raw data for machine learning models.

The process of converting raw data into informative signals a model can use.

Methods for estimating which input variables most influence a model's predictions.

The process of choosing the most useful input variables for a model.

A system for storing, sharing and serving machine learning features consistently.

Platforms that manage reusable features for training and production machine learning systems.

A training approach where models learn from data across devices or organizations without centralizing raw data.

An AI researcher known for major contributions to computer vision, ImageNet and human-centered AI.

The ability of a model to learn a task from only a small number of examples.

The process of adapting a pre-trained model to a narrower task, dataset or behavior.

A numerical method that breaks a physical system into smaller parts for simulation.

A formal logic system for representing objects, relationships and quantified statements.

A generative modeling approach that learns transformations from simple distributions to complex data.

Robot control focused on managing contact forces while interacting with objects or surfaces.

The market and business ecosystem built around large general-purpose AI models.

Large general-purpose models that can be adapted to many downstream tasks.

A mathematical transform that represents signals as combinations of frequencies.

Programmable hardware chips that can be configured for specialized computation.

Knowledge representation systems that organize information into structured frames and slots.

A branch of mathematics studying spaces of functions and transformations between them.

AI systems designed to play games, plan moves and learn strategies.

The study of strategic decisions among agents whose outcomes depend on each other.

Simulations that explore strategic behavior, incentives and outcomes among interacting agents.

Recurrent neural network units that use gates to manage information over sequences.

A probabilistic model that represents data as a mixture of several normal distributions.

The European Union regulation that sets rules for processing and protecting personal data.

Neural network architectures where a generator and discriminator learn through competition.

AI systems that create new text, images, audio, video, code or other content from learned patterns.

Models that learn a data distribution and use it to produce new examples or estimate likelihoods.

Optimization methods inspired by selection, mutation and recombination in evolution.

An open source Python library used for topic modeling, document similarity and vector representations of text.

A computer scientist known for foundational work in neural networks and deep learning.

Coordination between countries, organizations and institutions on shared AI risks, standards and benefits.

Generative pre-trained transformer models used for language generation, reasoning and multimodal AI tasks.

A visualization method that highlights image regions influencing a neural network's prediction.

Ensemble models that build strong predictors by adding many weak models step by step.

A training technique that limits gradient size to make neural network optimization more stable.

An optimization method that updates model parameters in the direction that reduces loss.

Neural networks designed to learn from graph-structured data such as nodes and relationships.

The mathematical study of nodes, edges and relationships between connected objects.

Parallel processors widely used to speed up AI training and inference.

A hyperparameter search method that tests combinations from a predefined parameter grid.

A distributed data processing ecosystem for storing and analyzing large datasets.

A visual representation that uses color intensity to show values, attention, density or importance.

Rule-of-thumb methods that find useful solutions when exhaustive search is too slow.

Probabilistic sequence models with hidden states that generate observable events.

Reinforcement learning that breaks complex decisions into smaller subgoals or policies.

Computing systems built for large-scale processing, simulation and scientific workloads.

A United States healthcare privacy and security law relevant to protected health information.

A chart that shows how numeric values are distributed across ranges or bins.

A computer vision feature descriptor based on gradients and edge directions in images.

Groups of high-performance computers connected to run demanding workloads in parallel.

An AI company and platform known for model hosting, datasets, libraries and community collaboration.

A widely used library for working with transformer models across text, vision and audio tasks.

Basic rights and freedoms that AI systems should respect across design, deployment and governance.

A design approach where people review, guide or correct AI system decisions.

The study and design of how people and robots communicate, cooperate and share spaces.

The process of finding model settings that improve performance without being learned directly from data.

The practical adjustment of model configuration values such as learning rate, depth or regularization.

An AI researcher known for work on deep learning, sequence models and large language models.

A computer vision and language task that generates text descriptions of images.

A computer vision task that assigns an image to one or more categories.

Computer vision methods that identify objects, scenes, people or patterns in images.

A computer vision task that divides an image into meaningful regions or objects.

Datasets where some classes or outcomes appear much more often than others.

The effect of AI on jobs, skills, roles, productivity and labor market change.

A model's ability to adapt to a task from examples or instructions inside the prompt.

Robots designed for manufacturing, logistics, inspection and other industrial tasks.

The process of using a trained model to produce predictions, classifications or generated outputs.

Runtime systems that execute trained models and serve their outputs to applications.

Techniques that make model inference faster, cheaper or more efficient.

Statistical methods used to draw conclusions about populations from samples.

The mathematical study of information, uncertainty, compression and communication.

Clear permission from people based on understandable information about data use, risks and choices.

The practice of managing infrastructure through versioned configuration files and automation.

A computer vision task that separates individual object instances at the pixel level.

Training that teaches a model to follow human instructions more reliably.

Legal rights that protect creations, inventions, brands, code, data or media.

Systems that perceive an environment, choose actions and pursue goals.

The degree to which people can understand how a model works or why it produced an output.

Security systems that detect suspicious activity, attacks or policy violations.

A method that infers reward functions by observing expert behavior.

A computer scientist who helped define the field of artificial intelligence and coined the term.

A probability distribution describing how multiple random variables behave together.

An AI researcher associated with work on recurrent neural networks, compression and general AI ideas.

Interactive documents that combine code, text, outputs and visualizations for analysis and experiments.

A clustering algorithm that groups data around a chosen number of centers.

A simple prediction method based on the closest examples in the training data.

An AI entrepreneur, investor and author associated with AI industry strategy and applications.

Recursive estimation methods used to track changing states under uncertainty.

A high-level deep learning API used to build and train neural networks.

The point after which a model may not have learned newer information from training data.

A technique where a smaller model learns from the outputs or behavior of a larger model.

Structured representations of entities and relationships used for search, reasoning and context.

Methods for encoding facts, rules, concepts and relationships so machines can use them.

An open source system for deploying, scaling and managing containerized applications.

A regularization method that encourages sparse model weights by penalizing absolute values.

A regularization method that discourages large weights by penalizing squared values.

A training technique that softens target labels to reduce overconfidence.

The task of predicting or generating text by learning patterns in language.

Models trained to process, predict or generate language.

A mathematical transform used to analyze signals, systems and differential equations.

Large neural language models trained on broad data and adapted for many text and reasoning tasks.

AI models that can process or generate multiple data types such as text, images, audio and video.

A probabilistic topic model that discovers themes in collections of documents.

A compressed representation space where models encode meaningful patterns or features.

A probability principle stating that averages stabilize as the number of observations grows.

A normalization method that stabilizes neural network activations within each example.

A chart that shows how model performance changes as training progresses or data grows.

A training hyperparameter that controls how large each parameter update is.

A schedule that gradually reduces the learning rate during training.

The planned adjustment of the learning rate over training time.

An NLP process that reduces words to their dictionary or base form.

A model explanation method that approximates complex predictions with simpler local explanations.

The branch of mathematics focused on vectors, matrices and linear transformations.

A statistical model that predicts a numeric value using a linear relationship between variables.

A classification loss metric that penalizes confident wrong probability predictions.

A programming approach based on facts, rules and logical inference.

A classification model that estimates the probability of a binary or categorical outcome.

Recurrent neural networks with gates designed to capture long-range sequence dependencies.

Models designed to handle unusually large prompts, documents or conversation histories.

A parameter-efficient fine-tuning method that adapts large models with low-rank update matrices.

The debated idea that machines could have or simulate conscious experience.

A field of AI where systems learn patterns from data instead of relying only on explicit rules.

Automatic translation of text or speech from one language to another using language models and statistical methods.

Probabilistic sequence models where the next state depends mainly on the current state.

A family of sampling methods that use Markov chains to approximate complex probability distributions.

Formal decision models where an agent chooses actions, receives rewards and moves between states.

Probabilistic models that describe systems through states and transition probabilities.

An early artificial intelligence researcher associated with cognitive models, robotics and symbolic AI.

Language models trained to predict hidden words or tokens from surrounding context.

The practice of keeping critical business data consistent across systems and teams.

A Python library for creating plots, charts and visualizations.

Mathematical methods that break a matrix into simpler components for analysis or computation.

A regression metric that averages absolute prediction errors.

A regression metric that averages squared prediction errors and penalizes large mistakes more strongly.

Research that tries to understand the internal mechanisms of neural networks.

Audio features that summarize the frequency characteristics of sound in a speech-friendly way.

Learning methods that help models adapt quickly to new tasks or learn how to learn.

The organization and governance of information that describes data, models, assets and workflows.

An architecture where applications are split into small services that communicate through APIs.

A generative AI image platform used to create images from text prompts and visual references.

A speculative idea of transferring or emulating a mind in a digital system.

An optimization method that updates model parameters using small batches of training examples.

A model architecture that routes inputs to selected specialized expert submodels.

Practices and tools for deploying, monitoring and maintaining machine learning systems in production.

Robots that can move through physical environments using sensors, planning and control.

Structured documents that describe a model's purpose, performance, limitations and intended use.

Techniques that make models smaller, faster or cheaper while trying to preserve useful quality.

An open protocol for connecting AI applications to tools, data sources and external context.

The process of moving a trained model into a usable production or application environment.

Changes in data or behavior that make a model less accurate or reliable over time.

Tools that help people inspect why models produce particular outputs or predictions.

The ongoing tracking of model quality, latency, data changes and production issues.

The infrastructure and runtime used to deliver model predictions to applications.

Attacks that try to copy a model's behavior or extract its value through queries or access.

An optimization technique that smooths parameter updates by accumulating past gradients.

The collection of metrics, events and logs to understand system health and behavior.

Computational methods that use repeated random sampling to estimate results.

Simulations that estimate possible outcomes by repeatedly sampling uncertain variables.

The study of how automated systems should handle decisions with moral consequences.

Computer vision methods that identify movement in video or sensor data.

The process of planning feasible movements for robots, vehicles or agents.

A decision problem that balances exploring new options with exploiting known rewards.

AI systems that work with several data types such as text, images, audio and video.

Methods for analyzing, searching and organizing music and audio information.

A deep learning framework historically used for building and training neural networks.

Sequences of neighboring words, characters or tokens used in language analysis.

A probabilistic classification method based on Bayes' theorem and independence assumptions.

A simple classifier that uses Naive Bayes probabilities to assign labels.

An NLP task that finds names of people, organizations, places and other entities in text.

The task of generating readable language from data, prompts or structured inputs.

The field of AI focused on analyzing, understanding and generating human language.

The field of AI focused on analyzing, understanding and generating human language.

NLP methods that extract meaning, intent, entities and relationships from language.

An optimization method that uses a look-ahead momentum step to improve gradient descent.

Simulations that model how systems, agents or nodes interact across networks.

Methods that automatically search for effective neural network designs.

Machine learning models built from connected layers that transform inputs into outputs.

Specialized hardware designed to accelerate neural network workloads.

Computing approaches inspired by the structure or behavior of biological nervous systems.

A philosopher known for work on existential risk, superintelligence and the long-term impact of AI.

A Python toolkit for natural language processing, text analysis and linguistic resources.

Techniques that remove unwanted noise from audio, images, signals or data.

Databases that store and query data outside the traditional relational table model.

A Python library for numerical computing with arrays, matrices and mathematical operations.

NVIDIA's parallel computing platform and programming model for GPU acceleration.

A computer vision task that finds and labels objects in images or video.

A computer vision task that follows objects across frames in video.

Machine learning that updates models continuously or incrementally as new data arrives.

Structured vocabularies that define concepts, relationships and rules in a domain.

An AI company and platform associated with large language models, APIs and AI products.

OpenAI's family of generative pre-trained transformer models.

Technology that converts text in images or scanned documents into machine-readable text.

The mathematical study of finding the best solution under goals and constraints.

Data that differs meaningfully from what a model saw during training or validation.

A model failure mode where it learns training data too closely and performs worse on new examples.

A Python library for working with tabular data, data frames and data analysis workflows.

The use of multiple processors or cores to perform computation at the same time.

The number of learned values inside a machine learning model.

Fine-tuning methods that adapt large models by training only a small number of parameters.

An NLP task that labels words by grammatical role such as noun, verb or adjective.

Visualizations that show how a feature affects model predictions on average.

Planning a route or sequence of movements through an environment.

Methods for detecting regularities, classes or structures in data.

The ability of AI systems to interpret sensory input such as images, sound or sensor data.

An early linear neural network model used for simple classification.

A language model metric that measures how well a model predicts text.

A computer scientist and AI educator known for work on AI, search and practical software systems.

The philosophical study of consciousness, mental states and the nature of mind.

The study of speech sounds and how they are produced, transmitted and perceived.

Structured workflows that connect steps such as data preparation, model execution and evaluation.

AI methods for choosing actions and arranging tasks over time under constraints.

A visualization library and platform for interactive charts and dashboards.

Reinforcement learning methods that optimize policies directly using gradient estimates.

Methods that improve an agent's decision policy in reinforcement learning.

Computer vision methods that estimate the position of body parts, joints or objects.

A technique that gives sequence models information about token order or position.

The study of how language meaning depends on context, intent and social use.

A classification metric measuring how many positive predictions were actually correct.

Building models that use data to forecast outcomes or estimate future behavior.

A probability assigned to a future event, class or outcome by a model.

Training that adjusts model behavior using human or synthetic preferences.

A dimensionality reduction method that finds major directions of variation in data.

Control over how personal or sensitive information is collected, used, shared, and protected.

Techniques for training or using models while limiting exposure of sensitive data.

Functions that describe how probability is distributed across continuous values.

The mathematical study of uncertainty, random events, and likelihood.

The AI capability of finding actions or decisions that move from a current state to a goal.

The practice of representing business, industrial, or computational processes so they can be analyzed or optimized.

A supervision approach that evaluates the steps a model takes rather than only its final answer.

A prompting pattern where several model calls are linked so each step feeds the next one.

The practice of designing instructions, context, and examples so an AI model produces better results.

An attack pattern where hidden or hostile instructions try to override the intended behavior of an AI system.

A platform or catalog where reusable prompts are published, sold, shared, or discovered.

Reusable prompt structures with placeholders for task-specific context or user input.

A formal logic system that works with statements that are either true or false.

A reinforcement learning algorithm that updates policies while limiting how far each update can move.

The process of removing less useful model weights, neurons, branches, or features to make a system smaller or faster.

An open-source machine learning framework widely used for building, training, and deploying neural networks.

A reinforcement learning method that learns the expected value of actions in different states.

A memory-efficient fine-tuning method that combines quantization with low-rank adaptation.

The process of representing model weights or activations with lower-precision numbers to reduce memory and compute costs.

A computing approach that uses quantum states such as superposition and entanglement to process information.

The study of using quantum computing ideas or hardware in machine learning workflows.

An NLP task where a system returns an answer to a question from text, data, or model knowledge.

AI systems designed to interpret questions and return direct answers from sources or learned representations.

A regression metric that estimates how much variance in the target is explained by a model.

An ensemble method that combines many decision trees to improve prediction stability and accuracy.

A hyperparameter tuning method that tests randomly selected combinations from a search space.

An inventor, author, and futurist known for work and predictions about AI, computing, and human-machine futures.

A branch of mathematics focused on real numbers, limits, continuity, sequences, and functions.

Systems that apply rules, logic, models, or search to draw conclusions and support decisions.

A metric that measures how many relevant items or positive cases a model successfully finds.

Systems that suggest items, content, people, or actions based on user behavior and other signals.

Neural networks designed to process sequences by carrying information from earlier steps to later ones.

A structured effort to test an AI system by actively looking for failures, misuse paths, and safety gaps.

A machine learning task that predicts a continuous numeric value rather than a class label.

A statistical method for studying relationships between variables and estimating numeric outcomes.

The work of meeting laws, standards, policies, and industry rules that apply to an AI system or data workflow.

A machine learning approach where an agent learns actions by receiving rewards or penalties from an environment.

A training method that uses human preference signals to guide model behavior.

The use of reward-based learning to train robots to act in physical or simulated environments.

The ability to repeat an experiment, model run, or analysis and obtain consistent results.

A set of practices for building AI systems that are safe, fair, transparent, accountable, and useful.

A pattern where a generative model retrieves external information before producing an answer.

A rule or model that assigns a score to behavior in reinforcement learning.

An optimization algorithm that adapts learning rates using a moving average of squared gradients.

The study of robot motion, positions, joints, and geometry without focusing on forces.

The use of cameras and computer vision so robots can perceive objects, spaces, and motion.

The ability of robots to move, grasp, place, or interact with physical objects.

A middleware ecosystem and set of tools for building modular robot software.

Transformer-based models adapted for robot perception, planning, control, or action prediction.

A model quality that describes how well an AI system keeps working under noise, shifts, attacks, or unusual inputs.

A chart that shows the tradeoff between true positive rate and false positive rate across classification thresholds.

A roboticist and AI researcher known for work on behavior-based robotics, autonomous robots, and embodied intelligence.

A regression metric that measures average prediction error with larger errors penalized more strongly.

AI systems that use explicit if-then rules or expert knowledge to make decisions.

Visual explanations that highlight the parts of an input that most influenced a model output.

An on-policy reinforcement learning algorithm that updates action values from the current state, action, reward, and next action.

Methods for supervising AI systems when direct human checking becomes too slow or too difficult.

Empirical relationships that describe how model performance changes with data, compute, and model size.

A chart that displays points on two axes to reveal relationships, clusters, and outliers.

An open-source Python library for classical machine learning, preprocessing, model selection, and evaluation.

An open-source Python ecosystem for scientific computing, optimization, statistics, and numerical methods.

A Python visualization library built on Matplotlib for statistical charts and exploratory data analysis.

Methods for exploring possible states, paths, or solutions to find a useful answer or plan.

Hardware designed to protect keys, data, execution, or model workloads against tampering and unauthorized access.

Cryptographic methods that let multiple parties compute a result without revealing their private inputs.

Communication and computation procedures designed to protect data, identity, integrity, and access.

A neural network mechanism that lets tokens in a sequence weigh the importance of other tokens.

Vehicles that use sensors, models, maps, and control systems to drive with limited or no human input.

A learning approach where models create training signals from the data itself instead of relying on manual labels.

Knowledge representations that connect concepts through labeled relationships.

The process of converting natural language into a structured meaning representation or executable query.

An NLP task that identifies who did what to whom, when, where, and how in a sentence.

A computer vision task that assigns a class label to each pixel in an image.

A vision and set of standards for making web data more structured, linked, and machine-readable.

A learning approach that combines a small amount of labeled data with a larger amount of unlabeled data.

Devices that measure signals from the physical world and provide data for AI or robotic systems.

A tokenizer and text segmentation system often used to prepare text for language models.

An NLP task that detects attitude, emotion, or opinion in text.

An NLP task that assigns a label to each token or step in a sequence.

A cloud model where code runs on managed infrastructure without the user managing servers directly.

GPU compute offered through managed, on-demand infrastructure without maintaining dedicated servers.

A branch of mathematics that studies collections of objects and operations on those collections.

Explanation scores that estimate how much each feature contributed to a model prediction.

Neural network architectures that compare two inputs using shared weights to measure similarity.

A classical computer vision feature detector and descriptor used to identify distinctive local image features.

A clustering metric that measures how well points fit within their own cluster compared with other clusters.

An optimization method inspired by cooling metals that explores solutions while gradually reducing randomness.

The philosophical idea that reality could be a computer simulation.

Methods for transferring policies, models, or skills learned in simulation to the real world.

Methods that let a robot build a map while estimating its own position in that map.

An oversampling technique that creates synthetic minority-class examples for imbalanced datasets.

The degree to which people and communities trust, understand, and are willing to use AI systems.

Computational modeling of people, groups, institutions, or societies and their interactions.

The current best-known level of performance or capability for a task, benchmark, or method.

An open-source NLP library for building production text processing pipelines in Python.

A distributed data processing engine used for large-scale analytics, machine learning, and data engineering.

A speech processing task that identifies which speaker is talking at each moment in an audio recording.

The analysis of audio through time-frequency visual representations of sound.

Techniques that improve audio quality by reducing noise, echo, distortion, or other unwanted signals.

The conversion of spoken audio into written text by an AI or signal processing system.

The generation of spoken audio from text, commands, or structured data.

Neural networks that communicate using short spike-like events rather than continuous activations.

A query language for selecting, filtering, joining, and managing data in relational databases.

A family of diffusion models used to generate or edit images from text prompts or reference inputs.

An ensemble learning method that combines several models through another model that learns how to use their outputs.

Pairs that describe a situation an agent is in and an action available or taken in that situation.

The use of statistical methods to describe data, test assumptions, and support decisions.

Mathematical descriptions of how values or outcomes are spread across possible cases.

The discipline of collecting, describing, modeling, and interpreting data under uncertainty.

An NLP preprocessing step that reduces words to a rough root form.

An optimization method that updates model parameters from small batches or samples of data.

Modeling that explicitly includes randomness, uncertainty, or probability in a system.

Mathematical models of systems that evolve over time with randomness.

Common words that are sometimes removed or downweighted during text processing.

The processing of data continuously as it arrives rather than after it is stored as a batch.

The idea of an AI system with broad, human-level understanding or general intelligence.

The distinction between hypothetical general intelligence and narrow systems built for specific tasks.

A computer scientist known for influential work in AI, rational agents, and AI safety.

A generative technique that applies the visual style of one image or domain to another piece of content.

A text segmentation approach that splits words into smaller reusable units for language models.

A machine learning approach that trains models from examples paired with known labels or target values.

Machine learning models that separate classes or fit boundaries using support vectors and margins.

A classical computer vision method for detecting and describing distinctive image features.

The monitoring of people, places, behavior, or systems, often using sensors, data, and automated analysis.

A set of methods inspired by the collective behavior of decentralized groups.

An AI approach that represents knowledge with explicit symbols, rules, and logic.

The creation of artificial data that resembles real data for training, testing, or privacy-preserving workflows.

AI-generated or AI-edited media such as images, video, speech, music, or text.

The modeling of complex systems with feedback loops, stocks, flows, and time-dependent behavior.

A high-priority instruction that defines how an AI assistant should behave inside an application.

A dimensionality reduction method often used to visualize high-dimensional data in two or three dimensions.

A training technique where a sequence model receives the correct previous output during training.

The idea of a future point where technological progress, often driven by AI, becomes extremely rapid or hard to predict.

Remote control of a robot, vehicle, or machine by a human operator.

A generation setting that controls how random or focused a language model output tends to be.

A reinforcement learning method that learns from differences between predicted and updated value estimates over time.

Specialized accelerator chips designed by Google for machine learning workloads.

An open-source machine learning framework used to build, train, and deploy models.

The process of checking whether a model, dataset, or AI system behaves correctly and meets requirements.

The task of producing a shorter version of a text while preserving its important information.

The generation of audio content from a text prompt or structured instruction.

The generation of images from natural language prompts or other conditioning inputs.

The conversion of written text into spoken audio.

The generation of video content from text prompts, images, or other conditioning inputs.

A weighting method that scores words by how important they are in a document relative to a collection.

An older Python library for numerical computation and deep learning research.

The analysis of data points ordered over time to find trends, patterns, and forecasts.

A computer scientist known for work on AI ethics, dataset documentation, bias, and accountability.

Machine learning designed to run on very small, low-power devices.

The maximum amount of text or context a model can process in one request or conversation window.

The process of splitting text into units that a language model can process.

The text units, such as words, subwords, or characters, that language models read and generate.

The ability of language models to call external tools, APIs, code, search, or databases while completing a task.

Methods for discovering themes or topics in a collection of documents.

A branch of mathematics that studies shapes, spaces, and properties preserved under continuous deformation.

The reuse of knowledge from one trained model or task to improve performance on another task.

Models based on attention mechanisms that are widely used for language, vision, audio, and multimodal AI.

Neural network architectures built around self-attention and feed-forward layers.

The practice of making AI system behavior, data use, limits, and decision processes understandable.

Rules or expectations that require disclosure about AI use, data, outputs, or system limitations.

A reinforcement learning algorithm that updates policies while constraining how much the policy can change.

AI that is designed and evaluated for reliability, safety, fairness, transparency, accountability, and usefulness.

The degree to which language model outputs are accurate, honest, and not misleading.

A proposed test of machine intelligence based on whether a machine can imitate human conversation convincingly.

A dimensionality reduction method used for visualization, clustering exploration, and manifold learning.

A modeling failure where a model is too simple or undertrained to capture important patterns in the data.

Aircraft that operate without an onboard human pilot, often using sensors, control systems, and autonomy.

Ground vehicles that operate without an onboard human driver.

A machine learning approach that finds patterns in data without labeled target answers.

A chart showing how model performance changes as a hyperparameter varies.

A function that estimates expected future reward from a state or state-action pair.

A database designed to store and search embeddings by similarity.

The tracking of model files, metadata, experiments, and changes over time.

Models that create, extend, or edit video from prompts, images, or other inputs.

Interactive simulated environments experienced through virtual reality interfaces.

The use of voice characteristics to identify or verify a person.

The generation of speech that imitates a particular person’s voice.

The transformation of one voice recording so it sounds like another voice or style.

Early training steps where the learning rate is gradually increased before the main schedule begins.

AI designed for specific tasks rather than broad general intelligence.

Vector representations of words that capture semantic or contextual relationships.

The task of choosing the intended meaning of a word in context.

A subword tokenization algorithm used in many language models.

Learned representations that help an AI system predict how an environment changes.

A computer scientist known for foundational work in deep learning and convolutional neural networks.

A computer scientist known for foundational work in deep learning and representation learning.

A security model that assumes no user, device, or network is trusted by default.

The ability of a model to perform a new task or recognize a new class without task-specific examples.