AI Glossary

Ablation tests the importance of a model component by removing or disabling it and measuring the resulting performance change.

Accuracy is the proportion of predictions a model classifies correctly out of all evaluated examples.

An activation function transforms a neuron’s weighted input and enables neural networks to learn nonlinear relationships.

Active learning is a machine learning approach that selects the most informative examples for human labeling to improve a model with less annotation work.

An adversarial attack uses carefully crafted input to mislead an AI model, bypass safeguards, expose data, or trigger an incorrect prediction.

Agentic AI describes systems that can plan, use tools, and carry out multiple actions with limited supervision to achieve a defined goal.

AGI is a proposed form of AI able to learn, reason, and perform intellectual tasks across many domains at human-level breadth.

An AI accelerator is specialized computing hardware designed to execute machine learning training or inference workloads efficiently.

AI alignment is the practice of directing AI systems toward intended goals and human values while keeping their behavior safe and controllable.

AI benchmarks are standardized tests and datasets used to compare model performance on defined capabilities, tasks, or safety criteria.

The study of principles and practices ensuring artificial intelligence is developed and used responsibly. It addresses fairness, transparency, accountability, and societal impact, helping build trust between humans and intelligent systems.

AI fluency is the practical ability to use, assess, and discuss AI systems with an informed understanding of their limitations.

The hardware, software, and cloud systems that power AI development, enabling data processing, model training, and large-scale deployment across industries.

AI safety is the field focused on preventing harmful, unreliable, or unintended behavior in artificial intelligence systems.

An AI workflow is the connected sequence of data, model, evaluation, deployment, and monitoring steps behind an AI application.

A set of rules or steps that guide computers in solving problems or completing tasks. In AI, algorithms process data, learn from patterns, and power everything from search engines to complex machine learning systems.

Anomaly detection is an AI technique that identifies unusual data, behavior, or events that may signal fraud, faults, errors, or security threats.

A structured way for software systems to communicate and exchange data. In AI, APIs connect models and applications, making it easier to integrate machine learning tools, automate workflows, and deliver intelligent features.

Artificial Intelligence is the science of building machines that can think, learn, and act like humans. It combines data, algorithms, and computational power to enable systems that understand language, recognize images, and make decisions across diverse industries - from healthcare and finance to education and entertainment.

ASI is a hypothetical form of artificial intelligence that would outperform humans across nearly every cognitive domain.

An attention mask tells a transformer which token positions it may consider and which positions should be ignored.

An attention mechanism helps an AI model focus on the most relevant parts of its input when interpreting information or generating an output.

An autoencoder is a neural network that compresses data into a latent representation and reconstructs it for denoising, learning, or anomaly detection.

Automation bias is the tendency to over-trust an automated recommendation even when contrary evidence or errors are visible.

AutoML automates parts of the machine learning workflow, including model selection, feature processing, and hyperparameter search.

An autoregressive model predicts each next value or token using values that appeared earlier in the sequence.

Backpropagation trains neural networks by tracing prediction errors backward through their layers and adjusting internal weights to improve future results.

A baseline is a simple reference result used to judge whether a more complex AI model provides meaningful improvement.

Batch size is the number of examples a machine learning model processes before updating its weights during training and gradient optimization.

A Bayesian network represents probabilistic relationships among variables using a directed graph and conditional probabilities.

Beam search is an AI decoding method that keeps several likely output sequences at each step to improve translation, transcription, or text generation.

BERT is an encoder-only transformer model that learns language representations by using context from both directions.

Bias is a systematic distortion in AI results that can produce unfair or inaccurate outcomes for particular groups, situations, or data patterns.

Massive, complex datasets that traditional tools cannot easily process. In AI, big data enables models to detect patterns, learn from vast examples, and make accurate predictions that drive smarter decisions and innovations.

A black box model is an AI system whose internal decision process is difficult to inspect or explain, even when its predictions are useful.

BLEU score evaluates generated text by measuring n-gram overlap with one or more human-written reference translations.

Causal language modeling trains a model to predict each next token without accessing future tokens in the sequence.

Chain-of-thought prompting encourages a model to work through intermediate reasoning steps before producing a final answer.

An AI-powered system that mimics human conversation through text or speech. Chatbots use natural language processing to assist users, automate support, and provide instant, personalized interactions across industries.

A checkpoint is a saved snapshot of model parameters and training state that can be restored or deployed later.

Class imbalance occurs when some target categories contain far fewer training or evaluation examples than others.

Classification is a machine learning task that assigns data to predefined categories based on patterns learned from labeled examples.

CLIP is a model approach that learns shared relationships between images and natural-language descriptions through contrastive training.

Closed source describes software or AI systems whose underlying code, weights, training details, or other components are restricted.

Clustering groups unlabeled data points so examples within a group are more similar than examples in other groups.

An AI field that aims to replicate human reasoning and learning. Cognitive computing systems analyze data, interpret context, and assist in complex decisions across sectors like healthcare, finance, and customer experience.

Compute is the processing capacity consumed when AI models are trained, evaluated, or used to generate predictions and outputs.

A branch of AI that allows computers to interpret and understand visual data such as images and videos. It powers applications from facial recognition and autonomous vehicles to medical imaging and augmented reality.

A confusion matrix counts correct and incorrect classification results for each actual and predicted class.

A context window is the amount of text and other tokenized information an AI model can consider during a single interaction.

Continual learning enables an AI model to learn from a changing stream of tasks or data without losing earlier knowledge.

Contrastive learning teaches a model useful representations by comparing examples that should be similar with examples that should differ.

A convolutional neural network is a deep learning model that detects spatial patterns in images and other grid-like data.

Cross-entropy measures the difference between a predicted probability distribution and the correct target distribution.

Cross-validation estimates model generalization by training and evaluating repeatedly on different partitions of the available data.

Data augmentation expands a training dataset by creating realistic variations of existing examples while preserving their intended labels.

Data labeling assigns categories, annotations, or target values to examples so supervised AI systems can learn and be evaluated.

The process of analyzing large datasets to uncover patterns, correlations, and insights. In AI, data mining supports model training, prediction, and optimization across industries like finance, healthcare, and marketing.

Data parallelism trains copies of the same model on different data batches and combines their parameter updates.

Data poisoning manipulates training data so an AI model learns harmful, biased, or attacker-chosen behavior.

An interdisciplinary field combining statistics, computing, and AI to extract insights from data. It drives innovation in predictive modeling, automation, and data-driven strategy across industries.

A decision tree predicts an outcome by repeatedly splitting data according to feature-based rules arranged in a branching structure.

A subset of machine learning that uses multi-layered neural networks to process data. It powers advanced AI applications such as speech recognition, autonomous driving, and generative models that simulate creativity.

Demographic parity is satisfied when a model produces a selected outcome at equal rates across defined demographic groups.

Differential privacy limits how much the presence of one person’s data can influence a released result or trained model.

A diffusion model generates new content by learning to transform random noise into structured images, audio, video, or other data.

Dimensionality reduction compresses data into fewer variables while preserving the structure or information most useful for analysis.

Double descent is a learning pattern where test error falls, rises near interpolation, and then falls again as model capacity increases.

Dropout regularizes a neural network by randomly disabling some units during training to reduce overreliance and overfitting.

An embedding is a numerical representation that places related words, images, documents, or other items close together in a mathematical space.

Unexpected or unprogrammed actions that arise as AI systems grow more complex. These behaviors can lead to surprising creativity or unpredictable outcomes, highlighting the importance of AI safety and alignment.

An encoder transforms raw input into a compact representation that preserves information useful for another model component or task.

End-to-end learning trains one model to map raw inputs directly to final outputs with minimal hand-built intermediate processing.

Ensemble methods combine predictions from multiple models to improve accuracy, stability, or robustness over a single model.

An epoch is one complete pass through the available training dataset during machine learning optimization.

Equalized odds requires a classifier to have matching true-positive and false-positive rates across protected groups.

An expert system uses encoded knowledge and inference rules to reproduce specialist decision-making within a narrow domain.

Explainable AI uses methods that help people understand, evaluate, and challenge how an AI system reaches a decision or prediction.

F1 score is the harmonic mean of precision and recall, balancing false positives and false negatives in one metric.

Factuality is the degree to which an AI system’s claims are consistent with accurate, verifiable information.

Feature engineering creates or transforms input variables so a machine learning model can learn useful patterns more effectively.

Feature extraction transforms raw input into informative measurements or representations that a machine learning model can use.

Federated learning trains a shared model across decentralized devices or organizations without centralizing their raw local data.

A feedforward neural network moves information from input to output through layers without recurrent feedback connections.

Few-shot learning allows an AI model to recognize a task or pattern from only a small number of examples.

Fine-tuning adapts a pretrained AI model to a specific task, domain, or behavior by continuing training on targeted examples.

Forward propagation passes input through a neural network layer by layer to calculate its prediction or generated output.

A foundation model is a large, broadly trained AI model that can be adapted or prompted for many different tasks and applications.

Function calling lets a language model request an approved external tool using structured arguments defined by an application.

A GAN trains a generator and discriminator in competition so the generator learns to create increasingly realistic synthetic data.

A powerful AI field that creates new content - from text and code to images and music. It learns from existing data to generate realistic, creative results that are reshaping industries and workflows.

GPT is a family of transformer language models pre-trained to predict text and adapted to generate useful responses from prompts.

A GPU is a highly parallel processor widely used to train and run neural networks and other compute-intensive AI workloads.

Gradient clipping limits the size of training gradients to prevent unstable parameter updates and exploding values.

Gradient descent is an optimization algorithm that reduces model error by repeatedly adjusting parameters in the direction indicated by the loss gradient.

A graph neural network (GNN) is a deep learning model that analyzes nodes and relationships in graph data for prediction and pattern discovery.

Grounding connects AI output to trusted documents, data, tools, or real-world context to improve relevance and reduce unsupported model responses.

Safety mechanisms and ethical constraints that guide AI systems to operate responsibly. They prevent harmful or biased outputs and ensure transparency, accountability, and alignment with human values.

When an AI model produces confident but incorrect or fabricated information. Hallucinations highlight the need for better data validation, model tuning, and safeguards to maintain trust and reliability.

A heuristic is a practical rule or shortcut that helps an AI system search, optimize, or make decisions without evaluating every possible option.

A hidden layer is an internal neural network layer that transforms inputs into representations used to produce the final output.

Human evaluation uses people to judge AI outputs on qualities that automated metrics cannot measure reliably.

Human-centered AI designs artificial intelligence around human goals, well-being, agency, accessibility, and accountable oversight.

HCI studies and designs the ways people interact with computers, software interfaces, and intelligent systems.

Human-in-the-loop (HITL) is an AI workflow where people review, guide, correct, or approve model decisions and sensitive automated actions.

Hybrid AI combines different artificial intelligence approaches, often linking machine learning with rules, search, or symbolic reasoning.

A preset configuration that determines how a machine learning model learns from data. Adjusting hyperparameters like learning rate and depth helps optimize performance and model accuracy.

Hyperparameter tuning searches for training and model settings that deliver strong performance on validation data.

An AI capability that enables computers to identify and classify objects within images. It’s used in facial recognition, manufacturing, healthcare, and security systems for automation and insight extraction.

In-context learning lets a language model infer a task from instructions or examples inside the prompt without changing the model’s trained parameters.

Inference is the process in which a trained AI model applies learned patterns to new input and produces a prediction, decision, or generated response.

Instance segmentation is a computer vision task that detects individual objects and assigns a precise pixel-level mask to every separate instance.

Instruction tuning trains an AI model on instruction-and-response examples so it follows natural-language tasks more reliably across different use cases.

Interpretability is the degree to which people can understand how a model uses inputs to produce a prediction or decision.

IoU measures how closely a predicted region overlaps the correct region by comparing their intersection with their union.

Jaccard similarity is a metric that measures overlap between two sets and is used in clustering, recommendations, text analysis, and image segmentation.

K-means clustering is an unsupervised learning algorithm that groups similar data points around k centroids without requiring labeled examples.

k-NN predicts a new example from the labels or values of its closest stored neighbors in feature space.

Knowledge distillation trains a smaller student model to imitate a larger teacher model, reducing AI inference cost, memory use, and latency.

A knowledge graph organizes entities and their relationships so AI systems can connect facts, answer questions, and retrieve structured context.

A KV cache stores attention keys and values from earlier tokens so a transformer can generate later tokens more efficiently.

A type of AI trained on massive text datasets to understand and generate human-like language. LLMs power chatbots, translation tools, and writing assistants, transforming communication and productivity.

Latency is the time an AI system takes to begin or finish responding after it receives an input.

Latent space is a learned numerical representation where an AI model organizes hidden features and relationships within images, text, or other data.

Learning rate is a training hyperparameter that controls how much an AI model changes its weights during each optimization update.

A type of AI that learns from recent experiences and data to make decisions. It’s used in technologies like autonomous driving, where context and real-time adaptation are essential.

Linear regression predicts a continuous value using a weighted linear relationship between input features and the target.

LLMOps is the discipline of developing, deploying, evaluating, monitoring, and maintaining applications built around large language models.

Logistic regression estimates the probability of a class by applying a logistic function to a weighted combination of features.

A loss function measures model error during training and gives an optimization algorithm the objective used to adjust AI model parameters.

Low-rank adaptation (LoRA) is a parameter-efficient fine-tuning method that trains small adapter weights instead of updating an entire AI model.

A branch of AI where computers learn from data to make predictions and improve performance over time. It underpins applications like fraud detection, recommendation engines, and predictive analytics.

A Markov chain models transitions between states where the next state depends only on the current state.

Masked language modeling trains a model to reconstruct hidden tokens using the surrounding context on both sides.

A mixture of experts (MoE) is an AI architecture that routes each input to selected subnetworks, increasing capacity without activating every parameter.

MLOps combines machine learning and software operations to deploy, monitor, update, and govern models reliably in production.

A model card is structured documentation describing an AI model’s intended uses, evaluation results, limitations, and risks.

Model Context Protocol (MCP) is an open standard that connects AI applications with external tools, data sources, and reusable context.

Model drift is the decline or change in AI performance that occurs when real-world data and relationships move away from training conditions.

Model parallelism divides one large model across multiple processors so its parameters and computation fit collectively.

Model quantization lowers the numerical precision of AI weights or calculations to reduce memory, inference time, energy use, and deployment cost.

Multi-head attention runs several attention operations in parallel so a model can capture different relationships in the same input.

Multimodal AI is artificial intelligence that understands or generates multiple data types, including text, images, audio, video, and sensor input.

Multitask learning trains one model on several related objectives so shared representations can improve multiple tasks.

Named entity recognition (NER) is an NLP task that finds and classifies people, places, organizations, dates, products, and other entities in text.

NLG is the branch of AI that produces human-readable language from data, prompts, representations, or structured plans.

An AI field focused on enabling machines to understand, interpret, and generate human language. NLP powers chatbots, voice assistants, translation tools, and sentiment analysis systems.

NLU enables computer systems to identify meaning, intent, entities, and relationships in human language.

Negative prompting tells a generative AI model what to avoid in an image or other output, helping steer style, content, and unwanted artifacts.

NeRF is a neural representation that learns a 3D scene from 2D images and renders it from new viewpoints.

An AI model inspired by the human brain that processes information through interconnected layers. Neural networks learn from data to recognize patterns, classify objects, and make intelligent decisions.

Neuro-symbolic AI combines neural learning with explicit symbols, rules, or logic to support structured reasoning.

Normalization transforms data or neural network activations into a consistent scale or distribution to support stable and efficient AI model training.

One-shot learning is an AI approach that recognizes a new task or category from one example by using knowledge learned during previous training.

An open-weight model makes its trained parameter files available for download, inspection, adaptation, or local deployment.

Optical character recognition (OCR) is an AI technology that converts text in images, scans, photos, and PDFs into searchable machine-readable data.

Optimization finds model parameters or decisions that best satisfy an objective while respecting practical resource and safety constraints.

A modeling issue where an AI system learns training data too precisely, reducing its ability to generalize. Managing overfitting ensures models perform reliably on new, unseen data.

Padding adds placeholder values so variable-length inputs can be processed together in a fixed-size machine learning batch.

A parameter is a learned numerical value, such as a neural network weight, that determines how an AI model transforms input into output.

An AI technique that detects regularities and relationships in data. It helps systems identify trends in speech, text, and images, forming the backbone of many intelligent applications.

A perceptron is a simple artificial neuron that combines weighted inputs to make a binary classification and forms a foundation of neural networks.

Perplexity measures how uncertain a language model is when predicting a sequence, with lower values indicating better prediction.

Precision is the proportion of predicted positive cases that are actually positive.

An AI-driven practice that analyzes historical data to forecast future outcomes. It helps businesses make proactive decisions, manage risk, and optimize performance through data insights.

An advanced AI approach that recommends actions based on predictive insights. It uses optimization and simulation to guide decision-making and improve operational outcomes.

Pretraining teaches an AI model broad patterns from large datasets before the model is prompted or adapted for specific downstream tasks.

An instruction or input given to an AI model that guides its response. Well-crafted prompts lead to more accurate, creative, and context-aware results in generative AI systems.

Prompt engineering is the practice of designing instructions, context, examples, and output rules that help generative AI perform tasks reliably.

Prompt injection is an AI security attack that uses malicious instructions to override a model’s rules, expose data, or trigger unauthorized tool actions.

Q-learning is a reinforcement learning algorithm that learns the long-term value of actions through rewards without requiring a model of the environment.

A revolutionary computing approach based on quantum mechanics that can process information at unprecedented speeds. It holds immense potential for accelerating AI training and complex problem-solving.

A reasoning model is optimized to spend additional computation on complex problems before returning an answer or action.

Recall is the proportion of all actual positive cases that a model successfully identifies.

A recurrent neural network (RNN) is a deep learning model that carries information across sequential inputs such as text, speech, and time-series data.

Regression is a supervised machine learning task that predicts continuous numerical values such as prices, demand, time, or risk.

Regularization discourages a model from fitting training data too narrowly, improving its ability to generalize to new examples.

A learning method where AI improves through trial and error, guided by rewards and penalties. It’s used in robotics, gaming, and autonomous systems to develop adaptive, goal-driven behavior.

RLHF trains an AI system using human judgments about which model outputs are more helpful, safe, or appropriate.

Retrieval-augmented generation (RAG) combines search with generative AI so responses use relevant documents, current data, and source context.

Scaling laws describe predictable relationships between model performance and factors such as compute, data, and parameter count.

Self-attention lets each element in a sequence weigh and combine information from other elements in that same sequence.

Self-supervised learning trains AI models with labels or objectives created from the data itself, reducing dependence on manual annotation.

Semantic search uses embeddings to retrieve information by meaning and context, finding relevant results even when documents use different words.

An AI method that detects emotions and opinions in text or speech. It’s used in marketing, social media monitoring, and customer feedback to measure public sentiment and brand perception.

Seq2Seq models transform one sequence into another, supporting translation, summarization, transcription, and similar tasks.

Softmax converts a set of numerical scores into a probability distribution whose values sum to one.

A Special Purpose Vehicle is a separate legal entity created to manage financial risk, hold assets, or fund specific projects. In AI, SPVs are used to support innovation, protect investors, and structure focused ventures such as model development or data infrastructure projects.

Organized, machine-readable information stored in formats like databases or spreadsheets. Structured data is key for efficient AI training, pattern discovery, and predictive modeling.

A learning approach where AI models train on labeled data with known outcomes. It powers tasks like classification, speech recognition, and predictive analytics across industries.

Symbolic AI represents knowledge with explicit symbols and rules that a system manipulates through logical reasoning.

Synthetic data is artificially generated information used to train, test, or evaluate AI systems when real data is limited, sensitive, or costly.

TAU-bench is a benchmark that tests how well AI agents interact with users and tools in realistic, multi-step scenarios, measuring not just success but reliability across repeated trials.

Temperature is a generative AI setting that adjusts how strongly a model favors likely tokens, influencing output consistency and variation.

A tensor is a multidimensional array of numbers used to represent and process data inside machine learning systems.

A basic text unit — such as a word, symbol, or character - that AI language models use to process and generate text. Tokens define how language models interpret and structure responses.

Tokenization splits text or other input into units that an AI model converts into identifiers and processes.

The dataset used to teach AI models how to perform tasks. It helps systems recognize patterns, make predictions, and improve performance through iterative learning.

A technique that adapts knowledge from one trained model to a new but related task. It speeds up training, improves efficiency, and reduces the need for large datasets.

A transformer is a neural network architecture built on attention and used in modern language, vision, audio, video, and multimodal AI models.

A benchmark test that measures a machine’s ability to mimic human intelligence. If an evaluator cannot distinguish the AI from a person, the system is said to have passed the test.

Data without a fixed format, such as text, images, or videos. AI tools use natural language processing and deep learning to extract meaning and insights from this type of data.

A method where AI models identify patterns in unlabeled data without predefined outputs. It’s used in clustering, anomaly detection, and exploratory data analysis.

A vanishing gradient occurs when training signals become extremely small as they move backward through a deep neural network.

A vector database stores embeddings and performs similarity search for semantic retrieval, recommendations, and retrieval-augmented generation.

A Vision Transformer applies transformer attention to image patches for visual classification, recognition, and representation learning.

An AI technology that converts spoken language into digital text or commands. It powers voice assistants, transcription tools, and hands-free control systems with growing accuracy.

Wake word detection is an AI speech technique that recognizes a short activation phrase while balancing missed commands, false triggers, and privacy.

A Wasserstein GAN (WGAN) is a generative model that uses a distribution-distance objective to improve training stability and synthetic output quality.

Weak AI is task-specific artificial intelligence designed for defined functions, unlike hypothetical general intelligence with broad human-like abilities.

A web crawler is an automated program that discovers and visits linked web pages for search indexing, monitoring, retrieval, or AI data collection.

Web scraping extracts structured data from webpages for analysis or AI workflows while requiring validation, provenance, privacy, and access compliance.

A weight is a learned numerical parameter that controls the influence of an input or connection within a machine learning model or neural network.

Weight decay is a regularization method that limits large model weights during training to reduce overfitting and improve generalization.

Weight initialization sets the starting parameters of a neural network so activations and gradients remain stable when model training begins.

A white box model is an AI system with transparent, inspectable decision logic that supports explanation, auditing, debugging, and human review.

Word error rate (WER) measures speech-recognition accuracy from word substitutions, deletions, and insertions compared with a reference transcript.

Word2Vec is a neural embedding method that represents words as vectors learned from context, enabling semantic similarity and language analysis.

Workflow automation coordinates tasks, data, approvals, and AI-assisted actions through a repeatable process with defined rules and oversight.

A world model is an AI representation that predicts how an environment may change, helping agents simulate outcomes before choosing an action.

An x-vector is a neural speaker embedding that represents a voice for speaker verification, diarization, clustering, and audio analysis.

Xavier initialization sets neural network weights using layer size to stabilize activations and gradients at the beginning of model training.

Xception is a computer vision architecture that uses depthwise separable convolutions for efficient image classification and feature extraction.

XGBoost is a gradient-boosting algorithm that combines decision trees for accurate classification, regression, ranking, and forecasting.

XLA (Accelerated Linear Algebra) is a compiler that optimizes tensor operations for faster and more efficient machine learning training and inference.

XLM-R is a multilingual transformer encoder that supports cross-lingual classification, entity recognition, and language understanding tasks.

The XOR problem is a classic machine learning example showing why neural networks need hidden layers and nonlinear activations for some patterns.

YAML is a human-readable configuration format used in machine learning pipelines, model experiments, deployments, and AI automation workflows.

YARN is Hadoop’s resource-management and scheduling layer for running distributed data processing, analytics, and machine learning workloads.

The Yeo-Johnson transformation reshapes numerical data, including zero and negative values, to stabilize variance and support machine learning models.

YOLO (You Only Look Once) is a real-time computer vision approach that detects and classifies multiple objects in one image processing pass.

A yottabyte is 10^24 bytes of digital information, a unit used to describe extremely large data volumes relevant to global computing and AI growth.

Youden's J statistic evaluates a binary classifier by combining sensitivity and specificity and can help select a decision threshold.

YUV color space separates image brightness from color information and is widely used in video compression, cameras, and computer vision pipelines.

Zero-shot learning lets an AI model perform a new task or recognize an unseen category without task-specific labeled examples.