# Glossary of AI, machine learning and data terms

485 terms used across this site, each defined in one line and linked to a full guide. Terms mentioned in articles link to the same guides.

Jump to a letter: [A](#a) · [B](#b) · [C](#c) · [D](#d) · [E](#e) · [F](#f) · [G](#g) · [H](#h) · [I](#i) · [J](#j) · [K](#k) · [L](#l) · [M](#m) · [N](#n) · [O](#o) · [P](#p) · [Q](#q) · [R](#r) · [S](#s) · [T](#t) · [U](#u) · [V](#v) · [W](#w) · [X](#x) · [Y](#y) · [Z](#z)

## A

**[A2A](/glossary/a2a)**: An open protocol, now under the Linux Foundation, with which agents discover each other and delegate tasks. [Full guide →](/tools/a2a)

**[A2C](/glossary/a2c)**: An actor-critic method where a critic estimates state value and the advantage reduces the variance of policy updates; updates run synchronously. [Full guide →](/tools/a2c)

**[A3C](/glossary/a3c)**: Runs many actor-critic workers in parallel that update shared parameters asynchronously. [Full guide →](/tools/a3c)

**[Activation functions](/glossary/activation-functions)**: An activation function is the non-linear function applied to a neuron's weighted sum before its output is passed on. [Full guide →](/tools/activation-functions)

**[Active learning](/glossary/active-learning)**: The model picks the most informative unlabelled examples for a human to label, reducing labelling cost. [Full guide →](/tools/active-learning)

**[Adapters](/glossary/adapters)**: Adds compact bottleneck layers between the frozen layers of a pre-trained model and trains only those. [Full guide →](/tools/adapters)

**[Advanced RAG](/glossary/advanced-rag)**: Extends basic retrieval with pre-retrieval steps such as query rewriting and post-retrieval steps such as reranking and compression. [Full guide →](/tools/advanced-rag)

**[Agent harness](/glossary/agent-harness)**: An agent harness is the software that wraps a language model and runs it as an agent, handling the loop, tools, context, permissions and limits. [Full guide →](/tools/agent-harness)

**[Agent identity and delegation](/glossary/agent-identity-and-delegation)**: Agent identity and delegation is the set of methods that give an AI agent its own identity and a limited, traceable right to act for a person. [Full guide →](/tools/agent-identity-and-delegation)

**[Agent memory](/glossary/agent-memory)**: Stores that let an agent recall facts, preferences and past work, from short-term context to long-term notes. [Full guide →](/tools/agent-memory)

**[Agentic commerce](/glossary/agentic-commerce)**: Agentic commerce is buying and selling in which an AI agent finds products, makes choices and completes payment on a person's behalf. [Full guide →](/tools/agentic-commerce)

**[Agentic RAG](/glossary/agentic-rag)**: An agent decides what to retrieve, from which source, and whether to search again before answering. [Full guide →](/tools/agentic-rag)

**[AI governance](/glossary/ai-governance)**: The policies, roles, processes and technical controls an organisation uses to manage AI risks across the lifecycle. [Full guide →](/tools/ai-governance)

**[AI watermarking and provenance](/glossary/ai-watermarking-and-provenance)**: Techniques that embed a hidden, detectable signal in AI-generated content, or attach signed records of its origin, so it can later be identified. [Full guide →](/tools/ai-watermarking-and-provenance)

**[Amazon Athena](/glossary/amazon-athena)**: An interactive query service that runs SQL against files in S3 using the Glue catalog, billed by data scanned. [Full guide →](/tools/amazon-athena)

**[Amazon EMR](/glossary/amazon-emr)**: A managed platform for running Spark, Hive, Flink and other open-source frameworks at scale. [Full guide →](/tools/amazon-emr)

**[Amazon Kinesis](/glossary/amazon-kinesis)**: AWS services for ingesting, storing and delivering streaming data in real time. [Full guide →](/tools/amazon-kinesis)

**[Amazon MSK](/glossary/amazon-msk)**: A managed service that runs Apache Kafka clusters with AWS security and operations. [Full guide →](/tools/amazon-msk)

**[Amazon RDS](/glossary/amazon-rds)**: A service that runs PostgreSQL, MySQL, Oracle, SQL Server and others with automated backups and failover. [Full guide →](/tools/amazon-rds)

**[Amazon Redshift](/glossary/amazon-redshift)**: A managed, massively parallel columnar data warehouse on AWS with provisioned and serverless options. [Full guide →](/tools/amazon-redshift)

**[Amazon S3](/glossary/amazon-s3)**: A highly durable object store that holds files of any size and format, organized into buckets. [Full guide →](/tools/amazon-s3)

**[Anomaly detection](/glossary/anomaly-detection)**: Identifying rows, events or sensor readings that deviate from normal behaviour. [Full guide →](/tools/anomaly-detection)

**[Apache Airflow](/glossary/apache-airflow)**: A platform that defines pipelines as code in directed graphs and schedules and monitors them. [Full guide →](/tools/apache-airflow)

**[Apache Arrow](/glossary/apache-arrow)**: A language-independent in-memory columnar format with libraries for zero-copy exchange, IPC and RPC (Flight). [Full guide →](/tools/apache-arrow)

**[Apache Beam](/glossary/apache-beam)**: A programming model and SDKs for data pipelines that run unchanged on several execution engines. [Full guide →](/tools/apache-beam)

**[Apache Flink](/glossary/apache-flink)**: A distributed engine for low-latency event-time stream processing with exactly-once state. [Full guide →](/tools/apache-flink)

**[Apache Hive](/glossary/apache-hive)**: A data warehouse system that runs SQL over files in distributed storage and keeps table metadata in the Hive Metastore. [Full guide →](/tools/apache-hive)

**[Apache Hudi](/glossary/apache-hudi)**: An open table format and data platform that supports record-level upserts, deletes and incremental queries on lake storage. [Full guide →](/tools/apache-hudi)

**[Apache Iceberg](/glossary/apache-iceberg)**: A table format with snapshots, hidden partitioning, schema evolution and ACID transactions over files in object storage. [Full guide →](/tools/apache-iceberg)

**[Apache Kafka](/glossary/apache-kafka)**: A durable, partitioned log that producers write to and consumers read from at high throughput. [Full guide →](/tools/apache-kafka)

**[Apache Parquet](/glossary/apache-parquet)**: An open, compressed columnar file format that stores data by column for efficient analytical reads. [Full guide →](/tools/apache-parquet)

**[Apache Pulsar](/glossary/apache-pulsar)**: A distributed pub-sub messaging and streaming system that separates brokers from storage (BookKeeper). [Full guide →](/tools/apache-pulsar)

**[Apache Spark](/glossary/apache-spark)**: A unified engine for large-scale batch, SQL, streaming and machine-learning workloads across a cluster. [Full guide →](/tools/apache-spark)

**[Architecture design](/glossary/architecture-design)**: Deciding layers, attention patterns, input encoding, heads and parameter budget for a model and its training objective. [Full guide →](/tools/architecture-design)

**[ArgoCD](/glossary/argocd)**: A controller that keeps a cluster in sync with configuration stored in Git. [Full guide →](/tools/argocd)

**[ARIMA](/glossary/arima)**: Combines autoregression, differencing and moving averages for stationary forecasting. [Full guide →](/tools/arima)

**[Attention mechanism](/glossary/attention-mechanism)**: A mechanism that computes weighted combinations of inputs, with weights based on relevance between queries and keys. [Full guide →](/tools/attention-mechanism)

**[Augmented reality](/glossary/augmented-reality)**: Technology that registers digital objects to the physical environment through a camera or headset. [Full guide →](/tools/augmented-reality)

**[Autoencoders](/glossary/autoencoders)**: A network that encodes an input into a smaller code and decodes it back, learning what matters. [Full guide →](/tools/autoencoders)

**[AWQ](/glossary/awq)**: Protects the small share of weights that matter most for activations before quantizing the rest to low precision. [Full guide →](/tools/awq)

**[AWS CDK](/glossary/aws-cdk)**: A framework that defines AWS infrastructure in TypeScript, Python and other languages and synthesizes CloudFormation. [Full guide →](/tools/aws-cdk)

**[AWS DataSync](/glossary/aws-datasync)**: A transfer service that copies data between on-premises storage and AWS, with scheduling, verification and filtering. [Full guide →](/tools/aws-datasync)

**[AWS Glue Data Catalog](/glossary/aws-glue-data-catalog)**: A managed metadata repository that stores table definitions, schemas and partitions, readable by Athena, EMR and other engines. [Full guide →](/tools/aws-glue-data-catalog)

**[AWS Lake Formation](/glossary/aws-lake-formation)**: A service that sets up lakes and applies fine-grained access control at database, table, column and row level on top of the Glue catalog. [Full guide →](/tools/aws-lake-formation)

**[Axial attention](/glossary/axial-attention)**: Applies attention along one axis at a time, which keeps cost manageable on two-dimensional data. [Full guide →](/tools/axial-attention)

**[Axolotl](/glossary/axolotl)**: A framework that fine-tunes language models from a YAML configuration across many techniques and models. [Full guide →](/tools/axolotl)

## B

**[Backpropagation](/glossary/backpropagation)**: An algorithm that uses the chain rule to compute gradients of the loss with respect to every weight. [Full guide →](/tools/backpropagation)

**[Bayesian deep learning](/glossary/bayesian-deep-learning)**: Combines deep networks with Bayesian methods to quantify what the model does not know. [Full guide →](/tools/bayesian-deep-learning)

**[Bayesian optimization](/glossary/bayesian-optimization)**: Builds a probabilistic surrogate of an objective and chooses the next point by balancing exploration and exploitation. [Full guide →](/tools/bayesian-optimization)

**[BentoML](/glossary/bentoml)**: A framework that packages models as services with dependencies and deploys them. [Full guide →](/tools/bentoml)

**[BERT](/glossary/bert)**: A transformer encoder pre-trained to predict masked words using context from both sides, then fine-tuned for understanding tasks. [Full guide →](/tools/bert)

**[Bias-variance trade-off](/glossary/bias-variance-trade-off)**: The idea that a model's error on new data comes from being too simple (bias), too sensitive to its training sample (variance), and unavoidable noise. [Full guide →](/tools/bias-variance-trade-off)

**[BigQuery](/glossary/bigquery)**: Google Cloud's fully managed, serverless data warehouse with separated storage and compute. [Full guide →](/tools/bigquery)

**[BitFit](/glossary/bitfit)**: Updates just the bias parameters of a model and keeps all other weights frozen. [Full guide →](/tools/bitfit)

**[bitsandbytes](/glossary/bitsandbytes)**: A library providing 8-bit optimizers and 4-bit and 8-bit quantized layers, used by QLoRA. [Full guide →](/tools/bitsandbytes)

**[BLIP-2](/glossary/blip-2)**: Connects a frozen image encoder to a frozen language model through a lightweight query transformer. [Full guide →](/tools/blip-2)

**[BM25](/glossary/bm25)**: A probabilistic ranking function that scores documents by term frequency and rarity. [Full guide →](/tools/bm25)

**[BPE](/glossary/bpe)**: A tokenization algorithm that repeatedly merges the most frequent adjacent symbol pairs into new tokens. [Full guide →](/tools/bpe)

**[BrainScaleS](/glossary/brainscales)**: A neuromorphic system that emulates neurons with analog circuits, running much faster than biological time. [Full guide →](/tools/brainscales)

**[Braintrust](/glossary/braintrust)**: A platform for running experiments, scoring outputs and tracking quality over time. [Full guide →](/tools/braintrust)

**[Bronze layer](/glossary/bronze-layer)**: The first medallion layer, which keeps source data as it arrived, with ingestion metadata. [Full guide →](/tools/bronze-layer)

**[Browser agents](/glossary/browser-agents)**: An agent drives a browser to read pages, fill forms and complete flows. [Full guide →](/tools/browser-agents)

**[BYOL](/glossary/byol)**: Trains an online network to predict the output of a slowly moving target network, with no negative examples. [Full guide →](/tools/byol)

## C

**[Camunda](/glossary/camunda)**: A platform that executes processes modelled in BPMN and DMN. [Full guide →](/tools/camunda)

**[Cassandra](/glossary/cassandra)**: A highly available database that scales writes across many nodes without a single point of failure. [Full guide →](/tools/cassandra)

**[CatBoost](/glossary/catboost)**: A boosting library that handles categorical variables with ordered target statistics. [Full guide →](/tools/catboost)

**[Causal inference](/glossary/causal-inference)**: Methods that estimate the effect of interventions rather than only associations. [Full guide →](/tools/causal-inference)

**[Change data capture](/glossary/change-data-capture)**: Reads a database's transaction log to emit every insert, update and delete as an event. [Full guide →](/tools/change-data-capture)

**[ChromaDB](/glossary/chromadb)**: A simple embedding store that is easy to embed in an application. [Full guide →](/tools/chromadb)

**[Chronos](/glossary/chronos)**: Amazon's family of models that tokenize time-series values and use language-model architectures to forecast. [Full guide →](/tools/chronos)

**[Circuit breaker](/glossary/circuit-breaker)**: Opens after repeated failures so a service fails fast and recovers, instead of piling up requests. [Full guide →](/tools/circuit-breaker)

**[Claude Agent SDK](/glossary/claude-agent-sdk)**: A Python and TypeScript SDK that exposes the same agent loop, tools and context management that power Claude Code. [Full guide →](/tools/claude-agent-sdk)

**[Claude Code](/glossary/claude-code)**: A tool that works in a terminal, desktop or browser, reads a codebase, edits files, runs commands and tests, and keeps working over many steps. [Full guide →](/tools/claude-code)

**[Claude Fable 5.1](/glossary/claude-fable-5-1)**: The Fable variant of Anthropic's most capable model tier, released with extra safety measures for biology, cybersecurity and research on language models. [Full guide →](/tools/claude-fable-5-1)

**[Claude Haiku 5.5](/glossary/claude-haiku-5-5)**: The smallest and fastest model in the current Claude family. [Full guide →](/tools/claude-haiku-5-5)

**[Claude Opus 5.5](/glossary/claude-opus-5-5)**: A flagship Claude model tuned for demanding reasoning, coding and multi-step agent work. [Full guide →](/tools/claude-opus-5-5)

**[Claude Sonnet 5.5](/glossary/claude-sonnet-5-5)**: A smart, efficient Claude model that balances quality, speed and cost. [Full guide →](/tools/claude-sonnet-5-5)

**[ClickHouse](/glossary/clickhouse)**: A very fast open-source database for real-time analytics over large event tables. [Full guide →](/tools/clickhouse)

**[CLIP](/glossary/clip)**: A model trained contrastively on image and caption pairs that embeds both into one space. [Full guide →](/tools/clip)

**[Cloud TPU](/glossary/cloud-tpu)**: Google's custom accelerators for machine-learning workloads. [Full guide →](/tools/cloud-tpu)

**[Clustering](/glossary/clustering)**: Partitions data into groups of similar points, as with k-means or density-based methods. [Full guide →](/tools/clustering)

**[CNN](/glossary/cnn)**: Networks that apply learned filters across local neighbourhoods, sharing weights over positions. [Full guide →](/tools/cnn)

**[ColBERT](/glossary/colbert)**: Stores an embedding per token and scores a document by matching query tokens against document tokens. [Full guide →](/tools/colbert)

**[Colossal-AI](/glossary/colossal-ai)**: A library that combines multiple parallelism strategies and memory optimizations for large models. [Full guide →](/tools/colossal-ai)

**[Computer use](/glossary/computer-use)**: The model sees the screen and clicks, types and scrolls to use applications. [Full guide →](/tools/computer-use)

**[Constitutional AI](/glossary/constitutional-ai)**: Has a model critique and revise its own answers against a written constitution, then trains on the improved answers and AI-labelled preferences. [Full guide →](/tools/constitutional-ai)

**[Context engineering](/glossary/context-engineering)**: The discipline of selecting, ordering and compressing instructions, data, tools and memory that enter a model's context window. [Full guide →](/tools/context-engineering)

**[Contextual Retrieval](/glossary/contextual-retrieval)**: Prepends a short, model-written description of where each chunk sits in its document before embedding and indexing. [Full guide →](/tools/contextual-retrieval)

**[Continued pre-training](/glossary/continued-pre-training)**: Keeps training a base model on a large, unlabelled domain corpus before any instruction tuning. [Full guide →](/tools/continued-pre-training)

**[Contrastive learning](/glossary/contrastive-learning)**: Trains an encoder so that two views of the same item land close in embedding space while different items land far apart. [Full guide →](/tools/contrastive-learning)

**[ControlNet](/glossary/controlnet)**: Adds spatial controls such as edges and depth to a pre-trained diffusion model. [Full guide →](/tools/controlnet)

**[ConvNeXt](/glossary/convnext)**: A pure convolutional network redesigned with transformer-era choices, competitive with vision transformers. [Full guide →](/tools/convnext)

**[Cost and latency budgets](/glossary/cost-and-latency-budgets)**: Explicit ceilings on tokens, tool calls and seconds for an agent run. [Full guide →](/tools/cost-and-latency-budgets)

**[CQRS](/glossary/cqrs)**: Command query responsibility segregation uses separate models for updating and querying data. [Full guide →](/tools/cqrs)

**[CrewAI](/glossary/crewai)**: A framework in which a crew of agents with roles and goals collaborate on tasks. [Full guide →](/tools/crewai)

**[Cross-encoder reranking](/glossary/cross-encoder-reranking)**: A model reads the query and each candidate together and produces a more accurate relevance score for the shortlist. [Full guide →](/tools/cross-encoder-reranking)

**[Cross-validation](/glossary/cross-validation)**: A method that rotates which part of the data is held out for testing, averaging results across folds. [Full guide →](/tools/cross-validation)

**[CUDA](/glossary/cuda)**: A programming model and toolkit for running general-purpose code on NVIDIA GPUs. [Full guide →](/tools/cuda)

**[Curriculum learning](/glossary/curriculum-learning)**: Orders or schedules training data by difficulty so the model learns simpler patterns first. [Full guide →](/tools/curriculum-learning)

**[Curriculum scheduling](/glossary/curriculum-scheduling)**: Plans when each kind of data enters training, usually moving from simple to complex. [Full guide →](/tools/curriculum-scheduling)

## D

**[Dagster](/glossary/dagster)**: A Python orchestrator where pipelines are defined as data assets with dependencies, lineage and checks. [Full guide →](/tools/dagster)

**[Data catalog](/glossary/data-catalog)**: A system of record for what data exists, who owns it and what it means. [Full guide →](/tools/data-catalog)

**[Data contract validation](/glossary/data-contract-validation)**: Automated checks that data meets an agreed schema, quality and freshness agreement between producers and consumers. [Full guide →](/tools/data-contract-validation)

**[Data lineage](/glossary/data-lineage)**: Records where data came from and which transformations it went through to reach a report or model. [Full guide →](/tools/data-lineage)

**[Data mesh](/glossary/data-mesh)**: An organisational and architectural approach where domain teams own data products on a shared self-serve platform. [Full guide →](/tools/data-mesh)

**[Data parallelism](/glossary/data-parallelism)**: Each device holds a full copy of the model, processes a slice of the batch, and gradients are averaged. [Full guide →](/tools/data-parallelism)

**[Data preprocessing](/glossary/data-preprocessing)**: The steps that turn raw, messy data into a clean, consistent numeric form that a model can be trained and evaluated on without leakage. [Full guide →](/tools/data-preprocessing)

**[Data synchronization](/glossary/data-synchronization)**: Techniques such as CDC, batch replication and reconciliation that keep two or more systems aligned. [Full guide →](/tools/data-synchronization)

**[Data tokenization](/glossary/data-tokenization)**: Swaps sensitive data such as card or national numbers for non-sensitive tokens, with the real values held in a protected vault or derived by a key. [Full guide →](/tools/data-tokenization)

**[Databricks](/glossary/databricks)**: A managed platform for data engineering, SQL analytics and machine learning on a lakehouse. [Full guide →](/tools/databricks)

**[dbt](/glossary/dbt)**: A tool that turns SQL select statements into versioned, tested and documented models in a warehouse. [Full guide →](/tools/dbt)

**[DDIM](/glossary/ddim)**: A sampler that skips many steps of the diffusion process while keeping sample quality. [Full guide →](/tools/ddim)

**[DDPG](/glossary/ddpg)**: An off-policy actor-critic algorithm that learns a deterministic policy over continuous action spaces. [Full guide →](/tools/ddpg)

**[DDPM](/glossary/ddpm)**: The foundational diffusion formulation that gradually adds noise and trains a network to reverse it. [Full guide →](/tools/ddpm)

**[Decision trees](/glossary/decision-trees)**: A model that splits data step by step on feature thresholds, forming a tree of decisions that ends in predictions. [Full guide →](/tools/decision-trees)

**[Deduplication](/glossary/deduplication)**: Detects exact and near duplicates in training data so the model does not memorize them and benchmarks stay clean. [Full guide →](/tools/deduplication)

**[DeepEval](/glossary/deepeval)**: An open-source framework that runs metric-based tests on model outputs inside a test suite. [Full guide →](/tools/deepeval)

**[DeepSeek V4.1](/glossary/deepseek-v4-1)**: The latest open-weight DeepSeek release, a mixture-of-experts model that is strong at reasoning and code. [Full guide →](/tools/deepseek-v4-1)

**[DeepSpeed ZeRO](/glossary/deepspeed-zero)**: A set of memory optimizations that partition training state across devices so much larger models fit. [Full guide →](/tools/deepspeed-zero)

**[Delta Lake](/glossary/delta-lake)**: A storage layer with a transaction log that adds ACID transactions, time travel and schema enforcement to files in a lake. [Full guide →](/tools/delta-lake)

**[Denoising autoencoders](/glossary/denoising-autoencoders)**: Trains a network to restore clean data from a corrupted copy, which yields robust representations. [Full guide →](/tools/denoising-autoencoders)

**[DETR](/glossary/detr)**: An end-to-end detector that treats detection as set prediction with a transformer. [Full guide →](/tools/detr)

**[Diffusion language models](/glossary/diffusion-language-models)**: A diffusion language model generates text by iteratively denoising a whole sequence, usually by filling in masked tokens over several steps, instead of predicting one token at a time from left to right. [Full guide →](/tools/diffusion-language-models)

**[Diffusion models](/glossary/diffusion-models)**: Generative models that learn to denoise data step by step, turning noise into a sample. [Full guide →](/tools/diffusion-models)

**[Digital twins](/glossary/digital-twins)**: Models that mirror the state and behaviour of a machine, process or network using live data. [Full guide →](/tools/digital-twins)

**[DINO](/glossary/dino)**: A student network learns to match a momentum teacher on different crops, producing strong features for vision transformers. [Full guide →](/tools/dino)

**[DiskANN](/glossary/diskann)**: A graph-based index designed to serve very large vector collections from disk with low memory. [Full guide →](/tools/diskann)

**[Docker](/glossary/docker)**: A tool that packages an application with its dependencies into a portable image. [Full guide →](/tools/docker)

**[Domain adaptation](/glossary/domain-adaptation)**: Techniques that align a model trained on a source distribution so it keeps working on a different target distribution. [Full guide →](/tools/domain-adaptation)

**[Domain-driven design](/glossary/domain-driven-design)**: A design approach that uses a shared language and bounded contexts to align code with the business. [Full guide →](/tools/domain-driven-design)

**[DoRA](/glossary/dora)**: Splits each weight into magnitude and direction and applies low-rank updates to the direction, which narrows the gap to full fine-tuning. [Full guide →](/tools/dora)

**[Double DQN](/glossary/double-dqn)**: Separates action selection from action evaluation to correct the overestimation bias of plain DQN. [Full guide →](/tools/double-dqn)

**[DPO](/glossary/dpo)**: Optimizes a model directly on preferred versus rejected answers, without training a separate reward model or running reinforcement learning. [Full guide →](/tools/dpo)

**[DQN](/glossary/dqn)**: A reinforcement-learning algorithm that approximates the action-value function with a deep network, using replay memory and a target network. [Full guide →](/tools/dqn)

**[Dropout](/glossary/dropout)**: A regularisation technique that zeroes random activations during training to prevent co-adaptation. [Full guide →](/tools/dropout)

**[DSPy](/glossary/dspy)**: A framework that expresses an LLM pipeline as modules and optimizes its prompts and weights against a metric. [Full guide →](/tools/dspy)

**[DuckDB](/glossary/duckdb)**: An embedded, columnar SQL engine that runs analytical queries inside your application or notebook. [Full guide →](/tools/duckdb)

**[Dueling DQN](/glossary/dueling-dqn)**: Splits the network into a state-value stream and an advantage stream, which helps when many actions have similar value. [Full guide →](/tools/dueling-dqn)

## E

**[E-commerce](/glossary/e-commerce)**: Platforms for catalogs, search, pricing, checkout and fulfilment. [Full guide →](/tools/e-commerce)

**[EfficientNet](/glossary/efficientnet)**: A family of CNNs scaled with a compound coefficient that balances depth, width and input resolution. [Full guide →](/tools/efficientnet)

**[Elasticsearch](/glossary/elasticsearch)**: A search engine that indexes documents for full-text, filtering and aggregation queries. [Full guide →](/tools/elasticsearch)

**[Embeddings](/glossary/embeddings)**: Dense numeric vectors produced by a model so that items with similar meaning are close together in vector space. [Full guide →](/tools/embeddings)

**[EMR on EKS](/glossary/emr-on-eks)**: Runs Spark jobs on an Amazon EKS cluster using the EMR runtime, so analytics shares infrastructure with other workloads. [Full guide →](/tools/emr-on-eks)

**[EMR Serverless](/glossary/emr-serverless)**: A deployment option that runs Spark and Hive jobs with automatic scaling and no cluster to provision. [Full guide →](/tools/emr-serverless)

**[Ensemble learning](/glossary/ensemble-learning)**: A family of methods that train several models and combine their predictions into one, usually more accurate and more stable, answer. [Full guide →](/tools/ensemble-learning)

**[EU AI Act](/glossary/eu-ai-act)**: An EU regulation that classifies AI systems by risk and sets obligations for providers and deployers, including general-purpose AI models. [Full guide →](/tools/eu-ai-act)

**[Evals](/glossary/evals)**: Test sets and graders that measure quality of a prompt, model or agent. [Full guide →](/tools/evals)

**[Evaluation metrics](/glossary/evaluation-metrics)**: Numbers that summarise how well a model's predictions match the truth, chosen to reflect what matters for the task. [Full guide →](/tools/evaluation-metrics)

**[Event sourcing](/glossary/event-sourcing)**: Persists every state change as an immutable event, from which current state is rebuilt. [Full guide →](/tools/event-sourcing)

**[Event-driven architecture](/glossary/event-driven-architecture)**: A style in which components publish and react to events asynchronously instead of calling each other directly. [Full guide →](/tools/event-driven-architecture)

**[EWC](/glossary/ewc)**: Protects parameters that were important for earlier tasks with a quadratic penalty weighted by their importance. [Full guide →](/tools/ewc)

**[ExLlamaV2](/glossary/exllamav2)**: An inference library with a flexible mixed-precision quantization format and optimized kernels. [Full guide →](/tools/exllamav2)

**[Experience replay](/glossary/experience-replay)**: Keeps a buffer of earlier examples and trains on them alongside the new task. [Full guide →](/tools/experience-replay)

## F

**[FAISS](/glossary/faiss)**: Meta's library for nearest-neighbour search and clustering of dense vectors, with CPU and GPU indexes. [Full guide →](/tools/faiss)

**[Feature engineering](/glossary/feature-engineering)**: The craft of creating, transforming and selecting input variables so that a model can learn the pattern more easily and more reliably. [Full guide →](/tools/feature-engineering)

**[Feature store](/glossary/feature-store)**: A system that defines, stores and serves ML features consistently for offline training and online inference. [Full guide →](/tools/feature-store)

**[Federated learning](/glossary/federated-learning)**: Each site trains locally and only model updates are aggregated centrally, so raw records never leave their owner. [Full guide →](/tools/federated-learning)

**[Few-shot learning](/glossary/few-shot-learning)**: Adapts a model to a new task using only a few labelled examples, often through in-context examples or small updates. [Full guide →](/tools/few-shot-learning)

**[Fine-tuning](/glossary/fine-tuning)**: Further training of a pretrained model on new data so it performs a specific task, domain or style better. [Full guide →](/tools/fine-tuning)

**[FinTech](/glossary/fintech)**: Software for payments, lending, banking and markets. [Full guide →](/tools/fintech)

**[FlashAttention](/glossary/flashattention)**: An attention kernel that computes exact attention in tiles to avoid reading and writing the large attention matrix to GPU memory. [Full guide →](/tools/flashattention)

**[Format-preserving encryption](/glossary/format-preserving-encryption)**: An encryption mode whose output has the same length and character set as the input, such as a number that stays a number. [Full guide →](/tools/format-preserving-encryption)

**[FSDP](/glossary/fsdp)**: PyTorch's native method for sharding parameters, gradients and optimizer states across workers. [Full guide →](/tools/fsdp)

**[FT-Transformer](/glossary/ft-transformer)**: Turns each feature into an embedding and processes them with a transformer. [Full guide →](/tools/ft-transformer)

**[Full fine-tuning](/glossary/full-fine-tuning)**: Continues training with all model parameters trainable, rather than a small add-on. [Full guide →](/tools/full-fine-tuning)

**[Function calling](/glossary/function-calling)**: A model returns a structured call to a named function with arguments, and the application runs it and returns the result. [Full guide →](/tools/function-calling)

## G

**[GaLore](/glossary/galore)**: Projects gradients into a low-rank space so optimizer states take far less memory, while training all weights. [Full guide →](/tools/galore)

**[GAN](/glossary/gan)**: Two networks compete: a generator produces samples and a discriminator tries to tell them from real data. [Full guide →](/tools/gan)

**[GAT](/glossary/gat)**: A graph network that learns attention weights over each node's neighbours. [Full guide →](/tools/gat)

**[Gaussian splatting](/glossary/gaussian-splatting)**: Represents a scene as many 3D Gaussians that render quickly with high quality. [Full guide →](/tools/gaussian-splatting)

**[GCN](/glossary/gcn)**: Aggregates information from a node's neighbours to learn node and graph representations. [Full guide →](/tools/gcn)

**[GDPR](/glossary/gdpr)**: The European law that governs the collection and processing of personal data. [Full guide →](/tools/gdpr)

**[Gemini 3.8 Flash](/glossary/gemini-3-8-flash)**: A latency-optimized member of the Gemini 3 generation. [Full guide →](/tools/gemini-3-8-flash)

**[Gemini 4](/glossary/gemini-4)**: Google's fourth-generation Gemini line, natively multimodal across text, images, audio and video. [Full guide →](/tools/gemini-4)

**[Gemma 4](/glossary/gemma-4)**: Google's open models derived from Gemini research, available in small sizes that run on modest hardware. [Full guide →](/tools/gemma-4)

**[Generative UI](/glossary/generative-ui)**: Generative UI is an approach in which a model decides, at run time, what interface to show the user, rather than relying only on fixed, pre-built screens. [Full guide →](/tools/generative-ui)

**[GGUF](/glossary/gguf)**: A single-file format that stores quantized weights and metadata for efficient local inference. [Full guide →](/tools/gguf)

**[GIS](/glossary/gis)**: Software and data for storing, analyzing and mapping geospatial information. [Full guide →](/tools/gis)

**[GitLab CI/CD](/glossary/gitlab-ci-cd)**: A pipeline system defined in a repository file that builds, tests and deploys code. [Full guide →](/tools/gitlab-ci-cd)

**[GLM-5.3](/glossary/glm-5-3)**: The latest GLM release from Z.ai, with strong coding and tool-use ability. [Full guide →](/tools/glm-5-3)

**[Gold layer](/glossary/gold-layer)**: The final layer, with aggregated and modelled tables shaped for reporting and applications. [Full guide →](/tools/gold-layer)

**[Google ADK](/glossary/google-adk)**: An open-source framework for building, evaluating and deploying agents, with support for multi-agent designs and Gemini. [Full guide →](/tools/google-adk)

**[GPT architecture](/glossary/gpt-architecture)**: A transformer that uses causal self-attention to predict the next token, trained on large text corpora and scaled up. [Full guide →](/tools/gpt-architecture)

**[GPT-5.6](/glossary/gpt-5-6)**: A late release in OpenAI's GPT-5 family. [Full guide →](/tools/gpt-5-6)

**[GPT-6](/glossary/gpt-6)**: OpenAI's sixth-generation model line, released in several tiers aimed at different cost and capability points. [Full guide →](/tools/gpt-6)

**[GPTQ](/glossary/gptq)**: Quantizes weights layer by layer to 4 bits or less, using approximate second-order information to limit error. [Full guide →](/tools/gptq)

**[Gradient boosting](/glossary/gradient-boosting)**: An ensemble method that adds decision trees one at a time, each fitted to the remaining errors of the current model. [Full guide →](/tools/gradient-boosting)

**[Gradient checkpointing](/glossary/gradient-checkpointing)**: Stores only some activations and recomputes the rest during the backward pass. [Full guide →](/tools/gradient-checkpointing)

**[Gradient descent](/glossary/gradient-descent)**: Gradient descent is an iterative method that lowers a loss by repeatedly moving parameters in the direction of steepest decrease. [Full guide →](/tools/gradient-descent)

**[Grafana](/glossary/grafana)**: A visualization platform that charts data from many sources. [Full guide →](/tools/grafana)

**[GraphQL](/glossary/graphql)**: An API style in which clients ask for exactly the fields they need in one request. [Full guide →](/tools/graphql)

**[GraphRAG](/glossary/graphrag)**: Builds a graph of entities and relations from the corpus and retrieves along its structure and community summaries. [Full guide →](/tools/graphrag)

**[GraphSAGE](/glossary/graphsage)**: Learns to sample and aggregate neighbour features, so it can embed nodes it never saw in training. [Full guide →](/tools/graphsage)

**[Grok 4.7](/glossary/grok-4-7)**: A model in xAI's Grok 4 line, with built-in tool use and live-data access. [Full guide →](/tools/grok-4-7)

**[gRPC](/glossary/grpc)**: A framework that uses Protocol Buffers and HTTP/2 for typed, efficient service calls. [Full guide →](/tools/grpc)

**[GRU](/glossary/gru)**: A recurrent unit with fewer gates than an LSTM, often similarly accurate and faster to train. [Full guide →](/tools/gru)

**[Guardrails](/glossary/guardrails)**: Checks on inputs, outputs and actions that block unsafe, off-topic or malformed behaviour. [Full guide →](/tools/guardrails)

**[Guardrails AI](/glossary/guardrails-ai)**: A framework that checks model responses against validators for structure, safety and content rules. [Full guide →](/tools/guardrails-ai)

## H

**[Hadoop HDFS](/glossary/hadoop-hdfs)**: An open-source framework for distributed storage (HDFS) and processing (MapReduce, YARN) on clusters of commodity servers. [Full guide →](/tools/hadoop-hdfs)

**[HashiCorp Vault](/glossary/hashicorp-vault)**: A service that stores secrets, issues short-lived credentials and offers encryption as a service. [Full guide →](/tools/hashicorp-vault)

**[Helm](/glossary/helm)**: A tool that templates and versions Kubernetes manifests as charts. [Full guide →](/tools/helm)

**[Hexagonal architecture](/glossary/hexagonal-architecture)**: Isolates business logic from infrastructure through explicit ports and adapters. [Full guide →](/tools/hexagonal-architecture)

**[HNSW](/glossary/hnsw)**: A layered proximity-graph index that gives fast and accurate approximate search. [Full guide →](/tools/hnsw)

**[Hugging Face](/glossary/hugging-face)**: A platform and set of libraries for models, datasets and training tools such as Transformers. [Full guide →](/tools/hugging-face)

**[Human-in-the-loop](/glossary/human-in-the-loop)**: Places human review at defined points of an automated workflow, especially before consequential actions. [Full guide →](/tools/human-in-the-loop)

**[Hybrid search](/glossary/hybrid-search)**: Merges lexical results such as BM25 with dense vector results and fuses their rankings. [Full guide →](/tools/hybrid-search)

**[HyDE](/glossary/hyde)**: Has a model write a hypothetical answer first and retrieves with its embedding instead of the raw query. [Full guide →](/tools/hyde)

**[Hyperparameter tuning](/glossary/hyperparameter-tuning)**: The search for the training settings, such as learning rate or tree depth, that give a model its best performance on unseen data. [Full guide →](/tools/hyperparameter-tuning)

## I

**[IA3](/glossary/ia3)**: Adapts a model by learning vectors that scale keys, values and feed-forward activations, adding very few parameters. [Full guide →](/tools/ia3)

**[IBM TrueNorth](/glossary/ibm-truenorth)**: IBM's neuromorphic processor built from many small neurosynaptic cores for low-power inference. [Full guide →](/tools/ibm-truenorth)

**[Iceberg cataloging](/glossary/iceberg-cataloging)**: Registers table metadata locations in a catalog such as Glue, Hive, REST or Nessie so that engines find and commit to tables consistently. [Full guide →](/tools/iceberg-cataloging)

**[Iceberg compaction](/glossary/iceberg-compaction)**: Maintenance that merges many small data files, applies deletes and expires old snapshots to keep queries fast. [Full guide →](/tools/iceberg-compaction)

**[Imputation networks](/glossary/imputation-networks)**: Models trained to predict missing cells from the observed cells of the same row and table. [Full guide →](/tools/imputation-networks)

**[In-context learning](/glossary/in-context-learning)**: The ability of a trained model to perform a new task from examples or instructions in its input, without changing its weights. [Full guide →](/tools/in-context-learning)

**[Industry 4.0](/glossary/industry-4-0)**: An approach that combines IoT, cloud, data and AI with automation to make manufacturing flexible and data-driven. [Full guide →](/tools/industry-4-0)

**[InfoNCE](/glossary/infonce)**: A loss that raises the score of a true pair relative to a set of negatives, a lower bound on mutual information. [Full guide →](/tools/infonce)

**[Informatica Axon](/glossary/informatica-axon)**: A governance product for business glossaries, ownership and policies. [Full guide →](/tools/informatica-axon)

**[Informatica EDC](/glossary/informatica-edc)**: A catalog that scans systems to discover, classify and map data assets and lineage. [Full guide →](/tools/informatica-edc)

**[Information theory](/glossary/information-theory)**: Entropy, mutual information and KL divergence, which underlie many training losses. [Full guide →](/tools/information-theory)

**[Infrastructure as code](/glossary/infrastructure-as-code)**: Describes servers, networks and services in files that are reviewed, tested and applied automatically. [Full guide →](/tools/infrastructure-as-code)

**[Intel Loihi](/glossary/intel-loihi)**: Intel's neuromorphic processor that runs spiking networks with event-driven, low-power computation. [Full guide →](/tools/intel-loihi)

**[IPO](/glossary/ipo)**: A preference-optimization objective that regularizes more strictly than DPO and avoids overfitting to preference labels. [Full guide →](/tools/ipo)

**[Isolation forest](/glossary/isolation-forest)**: Isolates points with random trees, so anomalies are separated in fewer splits. [Full guide →](/tools/isolation-forest)

**[IVF-PQ](/glossary/ivf-pq)**: Partitions vectors into clusters and compresses them with product quantization to save memory. [Full guide →](/tools/ivf-pq)

## J

**[JAX](/glossary/jax)**: A library for array computing with automatic differentiation, JIT compilation and easy parallelism on accelerators. [Full guide →](/tools/jax)

**[JWT](/glossary/jwt)**: A signed, compact token that carries claims between parties. [Full guide →](/tools/jwt)

## K

**[k-nearest neighbors](/glossary/k-nearest-neighbors)**: A method that predicts by finding the k closest training examples and using their labels or values. [Full guide →](/tools/k-nearest-neighbors)

**[Kafka Connect](/glossary/kafka-connect)**: A framework that moves data in and out of Kafka with configurable source and sink connectors. [Full guide →](/tools/kafka-connect)

**[Keras](/glossary/keras)**: A simple API for defining and training neural networks across several backends. [Full guide →](/tools/keras)

**[Kimi K3](/glossary/kimi-k3)**: A large open-weight model from Moonshot AI, built for long-context reasoning and agentic tasks. [Full guide →](/tools/kimi-k3)

**[Knowledge distillation](/glossary/knowledge-distillation)**: Training a compact model to match the outputs of a larger model, transferring much of its capability. [Full guide →](/tools/knowledge-distillation)

**[KServe](/glossary/kserve)**: A standard custom resource for serving models with autoscaling, canary rollouts and transformers. [Full guide →](/tools/kserve)

**[KTO](/glossary/kto)**: Aligns a model from simple good or bad labels on single answers, instead of paired comparisons. [Full guide →](/tools/kto)

**[Kubeflow](/glossary/kubeflow)**: A toolkit for pipelines, training operators and serving on Kubernetes. [Full guide →](/tools/kubeflow)

**[Kubernetes](/glossary/kubernetes)**: A system that deploys, scales and heals containerized applications across a cluster. [Full guide →](/tools/kubernetes)

**[KV cache](/glossary/kv-cache)**: Stored attention keys and values from previous tokens, reused so each new token only needs computation for itself. [Full guide →](/tools/kv-cache)

**[KVKK](/glossary/kvkk)**: Turkey's Personal Data Protection Law and the authority that enforces it. [Full guide →](/tools/kvkk)

## L

**[Lakehouse](/glossary/lakehouse)**: An architecture that adds transactions, schemas and performance to open data lake storage so it serves BI, data science and ML. [Full guide →](/tools/lakehouse)

**[Langfuse](/glossary/langfuse)**: A platform for tracing, prompt management, scoring and cost tracking of LLM applications. [Full guide →](/tools/langfuse)

**[LangGraph](/glossary/langgraph)**: A library for building stateful, long-running agents as graphs of nodes with checkpoints and human-in-the-loop steps. [Full guide →](/tools/langgraph)

**[LangSmith](/glossary/langsmith)**: A platform from LangChain for tracing runs, building datasets and evaluating prompts and agents. [Full guide →](/tools/langsmith)

**[Lava](/glossary/lava)**: A framework for developing neuromorphic applications that can target Loihi and conventional hardware. [Full guide →](/tools/lava)

**[Learning rate schedules](/glossary/learning-rate-schedules)**: A learning rate schedule is a rule that changes the optimiser's learning rate as training progresses. [Full guide →](/tools/learning-rate-schedules)

**[Learning to rank](/glossary/learning-to-rank)**: Learns a scoring function that orders items by relevance, using pointwise, pairwise or listwise losses. [Full guide →](/tools/learning-to-rank)

**[LightGBM](/glossary/lightgbm)**: A gradient-boosting framework that grows trees leaf-wise and uses histogram splits for speed. [Full guide →](/tools/lightgbm)

**[LIME](/glossary/lime)**: Explains one prediction by fitting a simple interpretable model around that instance. [Full guide →](/tools/lime)

**[Linear algebra](/glossary/linear-algebra)**: The mathematics of vectors and matrices that underlies neural networks and embeddings. [Full guide →](/tools/linear-algebra)

**[Linear regression](/glossary/linear-regression)**: A model that predicts a numeric target as a weighted sum of input features, fitted by minimising squared error. [Full guide →](/tools/linear-regression)

**[LLaMA-Factory](/glossary/llama-factory)**: An open-source toolkit that supports many models and tuning methods, with a command-line and web interface. [Full guide →](/tools/llama-factory)

**[llama.cpp](/glossary/llama-cpp)**: A library that runs quantized language models on CPUs and a range of GPUs. [Full guide →](/tools/llama-cpp)

**[LlamaIndex](/glossary/llamaindex)**: A framework for connecting language models to data through indexing, retrieval and agent workflows. [Full guide →](/tools/llamaindex)

**[LLaVA](/glossary/llava)**: Connects a vision encoder to a language model and fine-tunes it on visual instructions. [Full guide →](/tools/llava)

**[LLM routing](/glossary/llm-routing)**: LLM routing is the practice of choosing, for each incoming request, which language model or configuration should handle it. [Full guide →](/tools/llm-routing)

**[LLM-as-judge](/glossary/llm-as-judge)**: A model scores or compares responses against a rubric. [Full guide →](/tools/llm-as-judge)

**[lm-evaluation-harness](/glossary/lm-evaluation-harness)**: EleutherAI's framework that runs a large library of academic benchmarks against a model with a unified interface. [Full guide →](/tools/lm-evaluation-harness)

**[Logistic regression](/glossary/logistic-regression)**: A linear model that outputs class probabilities through the logistic (sigmoid) function. [Full guide →](/tools/logistic-regression)

**[Logistics](/glossary/logistics)**: The planning and execution of transport, warehousing and delivery. [Full guide →](/tools/logistics)

**[Long-running tasks](/glossary/long-running-tasks)**: Designs for agents that run unattended for long periods, with checkpoints, task lists and recovery. [Full guide →](/tools/long-running-tasks)

**[LoRA](/glossary/lora)**: Freezes the model and learns small low-rank update matrices for selected layers, so only a small fraction of parameters change. [Full guide →](/tools/lora)

**[Loss functions](/glossary/loss-functions)**: A loss function is a formula that scores how far a model's predictions are from the targets, giving training something to minimise. [Full guide →](/tools/loss-functions)

**[LSTM](/glossary/lstm)**: A recurrent network with gates that control what is stored, forgotten and emitted over time. [Full guide →](/tools/lstm)

**[LwF](/glossary/lwf)**: Keeps a new model's outputs on old tasks close to those of the previous model through distillation. [Full guide →](/tools/lwf)

## M

**[MAE](/glossary/mae)**: Masks a large part of the input and trains the model to reconstruct it, learning rich representations. [Full guide →](/tools/mae)

**[Mamba](/glossary/mamba)**: A sequence model whose state-space parameters depend on the input, giving linear-time processing of long sequences. [Full guide →](/tools/mamba)

**[Mask R-CNN](/glossary/mask-r-cnn)**: A two-stage detector that adds a pixel-level mask branch to Faster R-CNN. [Full guide →](/tools/mask-r-cnn)

**[Masked cell modeling](/glossary/masked-cell-modeling)**: The tabular counterpart of masked language modelling: random cells are masked and the model reconstructs them. [Full guide →](/tools/masked-cell-modeling)

**[Mastra](/glossary/mastra)**: A TypeScript framework for building agents, workflows and retrieval pipelines. [Full guide →](/tools/mastra)

**[MCP](/glossary/mcp)**: An open protocol, created by Anthropic, that lets a model use tools, data sources and prompts exposed by servers. [Full guide →](/tools/mcp)

**[Medallion architecture](/glossary/medallion-architecture)**: A lakehouse design that refines data in stages from raw to cleansed to business-ready. [Full guide →](/tools/medallion-architecture)

**[Megatron-LM](/glossary/megatron-lm)**: A library with tensor, pipeline and sequence parallelism that is tuned for NVIDIA GPU clusters. [Full guide →](/tools/megatron-lm)

**[Meta-learning](/glossary/meta-learning)**: Trains across many tasks so the model can adapt to a new one with very little data or few gradient steps. [Full guide →](/tools/meta-learning)

**[Microservices](/glossary/microservices)**: An architecture that splits a system into services that are deployed and scaled separately. [Full guide →](/tools/microservices)

**[Microsoft Agent Framework](/glossary/microsoft-agent-framework)**: A framework for building single and multi-agent applications in Python and .NET, generally available since April 2026. [Full guide →](/tools/microsoft-agent-framework)

**[MIDAS](/glossary/midas)**: An imputation method that trains an overcomplete denoising autoencoder and draws multiple completions. [Full guide →](/tools/midas)

**[Milvus](/glossary/milvus)**: A distributed vector database built for billion-scale similarity search. [Full guide →](/tools/milvus)

**[MIRAS](/glossary/miras)**: A research framework that describes memory modules by their attentional bias, retention gate and update rule. [Full guide →](/tools/miras)

**[MirrorMaker 2](/glossary/mirrormaker-2)**: Kafka's tool for mirroring topics and consumer offsets between clusters. [Full guide →](/tools/mirrormaker-2)

**[Mistral Large 4](/glossary/mistral-large-4)**: The largest model in Mistral's current line, strong at multilingual work and coding. [Full guide →](/tools/mistral-large-4)

**[Mixed precision](/glossary/mixed-precision)**: Uses 16-bit or 8-bit arithmetic for most operations and keeps sensitive values in 32-bit. [Full guide →](/tools/mixed-precision)

**[Mixture-of-Experts](/glossary/mixture-of-experts)**: A layer with many expert networks where a router activates only a few per token, growing capacity without growing compute. [Full guide →](/tools/mixture-of-experts)

**[MLflow](/glossary/mlflow)**: A tool for experiment tracking, model registry and deployment. [Full guide →](/tools/mlflow)

**[MoCo](/glossary/moco)**: Builds a large dictionary of negatives with a slowly updated momentum encoder and a queue. [Full guide →](/tools/moco)

**[Model cards](/glossary/model-cards)**: Structured documentation of a model's purpose, training data, evaluation, limitations and ethical considerations. [Full guide →](/tools/model-cards)

**[Model merging](/glossary/model-merging)**: Model merging creates a new model by combining the weights of two or more existing models, usually fine-tuned from the same base, without further gradient training. [Full guide →](/tools/model-merging)

**[Model monitoring](/glossary/model-monitoring)**: Continuous tracking of input data, predictions and outcomes of deployed models to detect drift and performance decay. [Full guide →](/tools/model-monitoring)

**[Modular RAG](/glossary/modular-rag)**: Treats retrieval, routing, memory and generation as modules that can be recombined for each use case. [Full guide →](/tools/modular-rag)

**[MongoDB](/glossary/mongodb)**: A database that stores flexible JSON-like documents. [Full guide →](/tools/mongodb)

**[MSK Replicator](/glossary/msk-replicator)**: An AWS feature that copies topics and consumer-group offsets between MSK clusters in the same or different regions. [Full guide →](/tools/msk-replicator)

**[mTLS](/glossary/mtls)**: TLS in which both client and server present certificates to prove their identity. [Full guide →](/tools/mtls)

**[Multi-agent RL](/glossary/multi-agent-rl)**: Reinforcement learning where multiple agents learn together, competing or cooperating, with shared or separate rewards. [Full guide →](/tools/multi-agent-rl)

**[Multi-agent systems](/glossary/multi-agent-systems)**: A design in which specialized agents with separate contexts divide work and exchange results. [Full guide →](/tools/multi-agent-systems)

**[Multimodal models](/glossary/multimodal-models)**: Models that take in or produce several data types, such as text, images, audio and video, within one network. [Full guide →](/tools/multimodal-models)

## N

**[N-BEATS](/glossary/n-beats)**: A fully connected architecture of stacked blocks that decompose a series into trend and seasonality. [Full guide →](/tools/n-beats)

**[n8n](/glossary/n8n)**: An automation platform with many integrations and nodes for AI steps. [Full guide →](/tools/n8n)

**[Naive Bayes](/glossary/naive-bayes)**: A probabilistic classifier that applies Bayes' theorem assuming features are independent given the class. [Full guide →](/tools/naive-bayes)

**[Naive RAG](/glossary/naive-rag)**: The basic pattern of embedding a question, fetching the nearest document chunks and giving them to a model as context. [Full guide →](/tools/naive-rag)

**[NATS](/glossary/nats)**: A fast messaging system with publish-subscribe, request-reply and a persistent layer called JetStream. [Full guide →](/tools/nats)

**[NCCL](/glossary/nccl)**: A library that implements fast all-reduce and other communication between GPUs. [Full guide →](/tools/nccl)

**[NeMo Guardrails](/glossary/nemo-guardrails)**: A toolkit that adds topic, safety and security rails to LLM applications using a dedicated rule language. [Full guide →](/tools/nemo-guardrails)

**[NeRF](/glossary/nerf)**: Represents a 3D scene as a neural network that renders new views from images. [Full guide →](/tools/nerf)

**[Neural networks](/glossary/neural-networks)**: A neural network is a model made of layers of simple units that each compute a weighted sum and a non-linear function, with the weights learned from data. [Full guide →](/tools/neural-networks)

**[Neural ODE](/glossary/neural-ode)**: Treats the depth of a network as continuous time and solves an ordinary differential equation to compute outputs. [Full guide →](/tools/neural-ode)

**[Normalization layers](/glossary/normalization-layers)**: Layers that rescale activations to a standard range, stabilising and accelerating training. [Full guide →](/tools/normalization-layers)

**[Normalizing flows](/glossary/normalizing-flows)**: Invertible transformations that map simple distributions to complex ones while keeping the density tractable. [Full guide →](/tools/normalizing-flows)

**[Norse](/glossary/norse)**: A PyTorch library for spiking neural networks and event-based learning. [Full guide →](/tools/norse)

**[NVIDIA B200](/glossary/nvidia-b200)**: A GPU with a new architecture, larger memory and support for lower-precision formats. [Full guide →](/tools/nvidia-b200)

**[NVIDIA H100](/glossary/nvidia-h100)**: A GPU with Transformer Engine and fast interconnect, widely used for training and inference. [Full guide →](/tools/nvidia-h100)

**[NVIDIA H200](/glossary/nvidia-h200)**: A Hopper GPU with 141 GB of HBM3e memory. [Full guide →](/tools/nvidia-h200)

## O

**[OAuth 2.0](/glossary/oauth-2-0)**: A framework in which an application obtains limited access to a resource on a user's behalf through tokens. [Full guide →](/tools/oauth-2-0)

**[OAuth for tool access](/glossary/oauth-for-tool-access)**: Uses OAuth flows so an agent acts on a user's behalf with scoped, revocable permissions. [Full guide →](/tools/oauth-for-tool-access)

**[OIDC](/glossary/oidc)**: An identity protocol on top of OAuth 2.0 that adds standard sign-in and user information. [Full guide →](/tools/oidc)

**[One-class SVM](/glossary/one-class-svm)**: A kernel method that fits a boundary around normal samples and flags points outside it. [Full guide →](/tools/one-class-svm)

**[Online travel](/glossary/online-travel)**: Platforms that aggregate availability and prices from many suppliers and let customers book in real time. [Full guide →](/tools/online-travel)

**[ONNX Runtime](/glossary/onnx-runtime)**: A runtime that executes models in the ONNX format across hardware backends. [Full guide →](/tools/onnx-runtime)

**[OpenAI Agents SDK](/glossary/openai-agents-sdk)**: A lightweight SDK with agents, handoffs, guardrails and tracing for multi-agent applications. [Full guide →](/tools/openai-agents-sdk)

**[OpenCV](/glossary/opencv)**: A library of image-processing and classical vision functions in C++ and Python. [Full guide →](/tools/opencv)

**[OpenTelemetry](/glossary/opentelemetry)**: APIs and SDKs for generating traces, metrics and logs. [Full guide →](/tools/opentelemetry)

**[Optimization](/glossary/optimization)**: The study of algorithms such as gradient descent that minimize objectives under constraints. [Full guide →](/tools/optimization)

**[Optimizers](/glossary/optimizers)**: Algorithms that update model parameters from gradients, such as SGD, momentum, Adam and AdamW. [Full guide →](/tools/optimizers)

**[Optuna](/glossary/optuna)**: A library that searches hyperparameters with efficient samplers and pruning. [Full guide →](/tools/optuna)

**[Oracle Database](/glossary/oracle-database)**: A commercial relational database widely used in large enterprises. [Full guide →](/tools/oracle-database)

**[Orchestrator-workers](/glossary/orchestrator-workers)**: One model breaks down a task dynamically and hands subtasks to workers, then combines their results. [Full guide →](/tools/orchestrator-workers)

**[ORPO](/glossary/orpo)**: Merges supervised fine-tuning and preference optimization into one stage with no reference model. [Full guide →](/tools/orpo)

**[OWASP](/glossary/owasp)**: A community that publishes widely used security guidance such as the OWASP Top 10. [Full guide →](/tools/owasp)

## P

**[PatchTST](/glossary/patchtst)**: Splits a series into patches and uses a transformer with channel independence. [Full guide →](/tools/patchtst)

**[PEFT](/glossary/peft)**: A library that implements LoRA, prefix tuning, prompt tuning and related methods on top of Hugging Face models. [Full guide →](/tools/peft)

**[Perceiver IO](/glossary/perceiver-io)**: A DeepMind architecture that cross-attends from a large input to a small learned latent array and decodes arbitrary outputs. [Full guide →](/tools/perceiver-io)

**[pgvector](/glossary/pgvector)**: A PostgreSQL extension that stores embeddings and supports exact and approximate nearest-neighbour search. [Full guide →](/tools/pgvector)

**[Phoenix](/glossary/phoenix)**: A tool for tracing LLM applications with OpenTelemetry and running evaluations on the traces. [Full guide →](/tools/phoenix)

**[Pinecone](/glossary/pinecone)**: A hosted vector database that stores embeddings and serves low-latency similarity search. [Full guide →](/tools/pinecone)

**[PINN](/glossary/pinn)**: A neural network trained with the governing physical equations as part of its loss, so predictions respect the physics. [Full guide →](/tools/pinn)

**[Pipeline parallelism](/glossary/pipeline-parallelism)**: Assigns consecutive groups of layers to different devices and streams micro-batches through them. [Full guide →](/tools/pipeline-parallelism)

**[Plan-and-execute](/glossary/plan-and-execute)**: A pattern where one step produces a plan and later steps execute and revise it. [Full guide →](/tools/plan-and-execute)

**[Polars](/glossary/polars)**: A Rust-based DataFrame library with lazy query optimisation and parallel execution on Arrow memory. [Full guide →](/tools/polars)

**[PostgreSQL](/glossary/postgresql)**: A mature SQL database with strong standards support, extensions and JSON capabilities. [Full guide →](/tools/postgresql)

**[PPO](/glossary/ppo)**: A policy-gradient method that clips the update size so each step stays close to the previous policy. [Full guide →](/tools/ppo)

**[Predictive maintenance](/glossary/predictive-maintenance)**: Uses sensor data and models to estimate failure risk and remaining useful life. [Full guide →](/tools/predictive-maintenance)

**[Prefect](/glossary/prefect)**: A Python workflow orchestrator where decorated functions become flows and tasks with retries, caching and monitoring. [Full guide →](/tools/prefect)

**[Prefix tuning](/glossary/prefix-tuning)**: Prepends trainable vectors to the attention keys and values at each layer while the model stays frozen. [Full guide →](/tools/prefix-tuning)

**[Presidio](/glossary/presidio)**: Microsoft's open-source toolkit for finding and anonymizing personally identifiable information in text and images. [Full guide →](/tools/presidio)

**[Principal component analysis](/glossary/principal-component-analysis)**: A linear method that finds orthogonal directions of maximum variance and projects data onto the top ones. [Full guide →](/tools/principal-component-analysis)

**[Probability](/glossary/probability)**: The framework for describing random variables, distributions and how likely outcomes are. [Full guide →](/tools/probability)

**[Prometheus](/glossary/prometheus)**: A time-series database and alerting system that scrapes metrics from services. [Full guide →](/tools/prometheus)

**[Prompt caching](/glossary/prompt-caching)**: Prompt caching reuses the already-computed internal state for an identical prompt prefix across requests, so only the new part of each prompt has to be processed. [Full guide →](/tools/prompt-caching)

**[Prompt engineering](/glossary/prompt-engineering)**: Techniques for phrasing instructions, examples, roles and output formats so a language model produces the intended result. [Full guide →](/tools/prompt-engineering)

**[Prompt tuning](/glossary/prompt-tuning)**: Trains a small set of embedding vectors that are prepended to the input, with the model frozen. [Full guide →](/tools/prompt-tuning)

**[Prompt-injection defense](/glossary/prompt-injection-defense)**: Techniques that stop text in web pages, files or tool results from hijacking the agent. [Full guide →](/tools/prompt-injection-defense)

**[PromptFoo](/glossary/promptfoo)**: A command-line tool that runs test cases and comparisons across prompts and models, including security tests. [Full guide →](/tools/promptfoo)

**[Prophet](/glossary/prophet)**: A forecasting model that fits trend, seasonality and holiday effects. [Full guide →](/tools/prophet)

**[Protegrity](/glossary/protegrity)**: A platform for tokenization, encryption and masking of sensitive data across systems. [Full guide →](/tools/protegrity)

**[Pruning](/glossary/pruning)**: Removing weights, channels or layers that contribute little, to shrink and speed up a network. [Full guide →](/tools/pruning)

**[PyCaret](/glossary/pycaret)**: A library that wraps preprocessing, model comparison and tuning into short commands. [Full guide →](/tools/pycaret)

**[Pydantic AI](/glossary/pydantic-ai)**: An agent framework from the Pydantic team that validates inputs and outputs with typed models. [Full guide →](/tools/pydantic-ai)

**[PySpark](/glossary/pyspark)**: The Python API of Apache Spark, covering DataFrames, SQL and MLlib. [Full guide →](/tools/pyspark)

**[PyTorch](/glossary/pytorch)**: A flexible framework with dynamic computation graphs and strong GPU support. [Full guide →](/tools/pytorch)

**[PyTorch Lightning](/glossary/pytorch-lightning)**: A library that organizes training loops, distributed training and logging. [Full guide →](/tools/pytorch-lightning)

## Q

**[Qdrant](/glossary/qdrant)**: A vector database with strong filtering, payloads and efficient indexing. [Full guide →](/tools/qdrant)

**[QLoRA](/glossary/qlora)**: Combines 4-bit quantization of the frozen base with LoRA adapters, which cuts memory enough to fine-tune large models on one GPU. [Full guide →](/tools/qlora)

**[Query rewriting](/glossary/query-rewriting)**: Rewrites, expands or decomposes a user query so that retrieval matches the relevant documents better. [Full guide →](/tools/query-rewriting)

**[Qwen 4](/glossary/qwen-4)**: The fourth generation of Alibaba's Qwen models, published in a range of sizes under open licences. [Full guide →](/tools/qwen-4)

**[Qwen3.8](/glossary/qwen3-8)**: A model in Alibaba's Qwen3 line with Flash variants aimed at speed. [Full guide →](/tools/qwen3-8)

## R

**[RabbitMQ](/glossary/rabbitmq)**: A broker that routes messages through exchanges and queues with flexible delivery patterns. [Full guide →](/tools/rabbitmq)

**[RAGAS](/glossary/ragas)**: A library that scores faithfulness, answer relevance, context precision and recall for retrieval-augmented systems. [Full guide →](/tools/ragas)

**[Rainbow](/glossary/rainbow)**: Combines prioritized replay, multi-step returns, distributional values, noisy networks, double and dueling learning in a single agent. [Full guide →](/tools/rainbow)

**[Random forest](/glossary/random-forest)**: Averages many decorrelated trees, each trained on a random sample of rows and features. [Full guide →](/tools/random-forest)

**[Ray Serve](/glossary/ray-serve)**: A serving library built on Ray that composes models and scales them across a cluster. [Full guide →](/tools/ray-serve)

**[Ray Tune](/glossary/ray-tune)**: A library that runs many trials in parallel across a cluster with early stopping. [Full guide →](/tools/ray-tune)

**[ReAct loop](/glossary/react-loop)**: An agent pattern in which the model writes a thought, takes an action with a tool, observes the result and repeats. [Full guide →](/tools/react-loop)

**[Real-time systems](/glossary/real-time-systems)**: Systems that must process input and respond within bounded time. [Full guide →](/tools/real-time-systems)

**[Reasoning models](/glossary/reasoning-models)**: Language models trained to produce long internal reasoning before the final answer, trading extra inference time for accuracy. [Full guide →](/tools/reasoning-models)

**[Recommendation](/glossary/recommendation)**: Predicts which items a user will like from interactions and attributes. [Full guide →](/tools/recommendation)

**[Recurrent neural networks](/glossary/recurrent-neural-networks)**: A neural network that processes a sequence one step at a time, updating a hidden state that summarises everything seen so far. [Full guide →](/tools/recurrent-neural-networks)

**[Red-teaming](/glossary/red-teaming)**: Systematically testing a model or application with adversarial inputs to find unsafe or unintended behaviour. [Full guide →](/tools/red-teaming)

**[Redis](/glossary/redis)**: A very fast key-value store with data structures, caching and pub-sub. [Full guide →](/tools/redis)

**[Reflection](/glossary/reflection)**: A pattern where the model reviews its output, finds mistakes and revises it before finishing. [Full guide →](/tools/reflection)

**[Regression boards](/glossary/regression-boards)**: A standing set of benchmark rows that is re-run in full after each change so that a new fix cannot silently break an old result. [Full guide →](/tools/regression-boards)

**[Regularization](/glossary/regularization)**: Any technique that limits or guides what a model learns so that it performs better on unseen data, not just on its training set. [Full guide →](/tools/regularization)

**[REINFORCE](/glossary/reinforce)**: A Monte Carlo policy-gradient algorithm that increases the probability of actions in proportion to the return that followed them. [Full guide →](/tools/reinforce)

**[Reinforcement learning](/glossary/reinforcement-learning)**: An agent acts in an environment, receives rewards, and learns a policy that maximizes long-term reward. [Full guide →](/tools/reinforcement-learning)

**[ResNet](/glossary/resnet)**: A convolutional architecture whose residual connections make very deep networks trainable. [Full guide →](/tools/resnet)

**[RLAIF](/glossary/rlaif)**: Uses an AI model, rather than humans, to produce the preference labels used for alignment. [Full guide →](/tools/rlaif)

**[RLHF](/glossary/rlhf)**: Trains a reward model from human preference comparisons and then optimizes the language model against it with reinforcement learning. [Full guide →](/tools/rlhf)

**[rsLoRA](/glossary/rslora)**: Changes the LoRA scaling factor so that higher adapter ranks train stably and keep improving. [Full guide →](/tools/rslora)

## S

**[S4](/glossary/s4)**: A state-space model with structured parameters that handles very long sequences efficiently. [Full guide →](/tools/s4)

**[SAC](/glossary/sac)**: An off-policy actor-critic algorithm that maximizes reward plus policy entropy, which encourages exploration and stability. [Full guide →](/tools/sac)

**[Saga pattern](/glossary/saga-pattern)**: Splits a business transaction into local steps, each with a compensating action if later steps fail. [Full guide →](/tools/saga-pattern)

**[SageMaker](/glossary/sagemaker)**: A managed service for building, training and deploying models on AWS. [Full guide →](/tools/sagemaker)

**[SAINT](/glossary/saint)**: A transformer that attends across features and across rows, with contrastive pre-training. [Full guide →](/tools/saint)

**[SAM](/glossary/sam)**: A promptable segmentation foundation model from Meta. [Full guide →](/tools/sam)

**[Sandboxing](/glossary/sandboxing)**: Executes code and tools inside a restricted container or virtual machine. [Full guide →](/tools/sandboxing)

**[ScaNN](/glossary/scann)**: A library that uses anisotropic quantization to give accurate maximum-inner-product search. [Full guide →](/tools/scann)

**[Schema evolution](/glossary/schema-evolution)**: Adding, renaming, dropping or widening columns without rewriting data or breaking readers. [Full guide →](/tools/schema-evolution)

**[Schema Registry](/glossary/schema-registry)**: A service that stores Avro, Protobuf and JSON schemas and checks that producers and consumers stay compatible. [Full guide →](/tools/schema-registry)

**[Schema-drift detection](/glossary/schema-drift-detection)**: Compares incoming data's structure against its expected schema and alerts or adapts when columns or types change. [Full guide →](/tools/schema-drift-detection)

**[scikit-learn](/glossary/scikit-learn)**: A library of consistent APIs for preprocessing, models, pipelines and evaluation. [Full guide →](/tools/scikit-learn)

**[Self-supervised learning](/glossary/self-supervised-learning)**: Creates its own training signal by masking, corrupting or transforming the input and asking the model to recover or relate it. [Full guide →](/tools/self-supervised-learning)

**[Semantic sector heads](/glossary/semantic-sector-heads)**: Classification heads, with a coarse-to-fine hierarchy, that map a table's content to an industry sector. [Full guide →](/tools/semantic-sector-heads)

**[Semi-supervised learning](/glossary/semi-supervised-learning)**: Training with a small labelled set and a much larger unlabelled set, so that the unlabelled data shapes the decision boundary. [Full guide →](/tools/semi-supervised-learning)

**[SentencePiece](/glossary/sentencepiece)**: A library that trains subword tokenizers directly from raw text, without pre-splitting into words. [Full guide →](/tools/sentencepiece)

**[SFT](/glossary/sft)**: Trains a model on curated prompts and ideal responses so it follows instructions in the desired style. [Full guide →](/tools/sft)

**[SGLang](/glossary/sglang)**: A serving framework with efficient prefix caching and a language for structured generation. [Full guide →](/tools/sglang)

**[SHAP](/glossary/shap)**: Attributes a prediction to its input features using Shapley values from game theory. [Full guide →](/tools/shap)

**[Silver layer](/glossary/silver-layer)**: The second layer, where data is validated, deduplicated, typed and joined into consistent entities. [Full guide →](/tools/silver-layer)

**[SimCLR](/glossary/simclr)**: A self-supervised framework that maximizes agreement between two augmented views of the same image using a contrastive loss. [Full guide →](/tools/simclr)

**[SimPO](/glossary/simpo)**: A reference-free preference objective that uses the average log-probability of a response as an implicit reward. [Full guide →](/tools/simpo)

**[Single message transforms](/glossary/single-message-transforms)**: Lightweight transformations applied to each record as it passes through a connector, such as masking, renaming or routing. [Full guide →](/tools/single-message-transforms)

**[Skills](/glossary/skills)**: Folders of instructions, scripts and resources that an agent loads when a task needs them. [Full guide →](/tools/skills)

**[Small language models](/glossary/small-language-models)**: Language models with relatively few parameters that run on modest hardware or devices, often fine-tuned or distilled for specific tasks. [Full guide →](/tools/small-language-models)

**[Smart city](/glossary/smart-city)**: Applies sensing, data platforms and analytics to services such as transport, energy, planning and public safety. [Full guide →](/tools/smart-city)

**[Smart manufacturing](/glossary/smart-manufacturing)**: Connects machines, sensors and production systems so that quality, maintenance and scheduling can be monitored and optimized with data and models. [Full guide →](/tools/smart-manufacturing)

**[smolagents](/glossary/smolagents)**: A small library where agents write and run code actions to solve tasks. [Full guide →](/tools/smolagents)

**[SmoothQuant](/glossary/smoothquant)**: Moves quantization difficulty from activations to weights with a mathematically equivalent scaling, enabling 8-bit weights and activations. [Full guide →](/tools/smoothquant)

**[snnTorch](/glossary/snntorch)**: A PyTorch library for building and training spiking neural networks with surrogate gradients. [Full guide →](/tools/snntorch)

**[Snowflake](/glossary/snowflake)**: A managed platform that separates storage from compute and runs SQL analytics at scale. [Full guide →](/tools/snowflake)

**[Spark Streaming](/glossary/spark-streaming)**: Processes live data in micro-batches or continuous mode with the same API used for batch. [Full guide →](/tools/spark-streaming)

**[Spark UDF de-identification](/glossary/spark-udf-de-identification)**: Applies user-defined functions in Spark to mask, tokenize or encrypt fields as part of a pipeline. [Full guide →](/tools/spark-udf-de-identification)

**[Speculative decoding](/glossary/speculative-decoding)**: An inference technique where a cheap draft proposes tokens and the main model verifies several at once, preserving its output distribution. [Full guide →](/tools/speculative-decoding)

**[Spiking neural networks](/glossary/spiking-neural-networks)**: Networks whose neurons fire timed spikes, mimicking biological neurons and enabling event-driven computation. [Full guide →](/tools/spiking-neural-networks)

**[SpikingJelly](/glossary/spikingjelly)**: A PyTorch-based framework for spiking networks, datasets and neuromorphic training. [Full guide →](/tools/spikingjelly)

**[SpiNNaker](/glossary/spinnaker)**: A massively parallel computer built from many small cores to simulate spiking neural networks in real time. [Full guide →](/tools/spinnaker)

**[Stable Diffusion](/glossary/stable-diffusion)**: A latent diffusion model that generates images from text prompts. [Full guide →](/tools/stable-diffusion)

**[Statistics](/glossary/statistics)**: Methods for estimation, testing and uncertainty quantification from samples. [Full guide →](/tools/statistics)

**[Structured outputs](/glossary/structured-outputs)**: Constrains generation to a JSON schema or type so that results are always parseable. [Full guide →](/tools/structured-outputs)

**[StyleGAN](/glossary/stylegan)**: A GAN architecture with a style-based generator that controls features at different scales. [Full guide →](/tools/stylegan)

**[Sub-agents](/glossary/sub-agents)**: An agent spawns focused helpers whose work returns as a short result, which keeps the main context clean. [Full guide →](/tools/sub-agents)

**[Supervised learning](/glossary/supervised-learning)**: A learning setup where every training example comes with the correct answer, and the model learns the mapping from inputs to that answer. [Full guide →](/tools/supervised-learning)

**[SVM](/glossary/svm)**: A model that finds the maximum-margin boundary, optionally in a kernel-induced feature space. [Full guide →](/tools/svm)

**[Swin](/glossary/swin)**: A vision transformer that computes attention inside local windows that shift between layers, giving a feature hierarchy. [Full guide →](/tools/swin)

**[Synaptic Intelligence](/glossary/synaptic-intelligence)**: Accumulates how much each weight contributed to loss reduction during training and penalizes changes to important ones. [Full guide →](/tools/synaptic-intelligence)

**[Synthetic data](/glossary/synthetic-data)**: Data produced by a model, program or simulator rather than recorded from real events, used to train, test or share machine learning systems. [Full guide →](/tools/synthetic-data)

**[Synthetic data generation](/glossary/synthetic-data-generation)**: Produces artificial datasets from generators, priors or models to cover patterns that real data lacks. [Full guide →](/tools/synthetic-data-generation)

## T

**[t-SNE](/glossary/t-sne)**: A non-linear method that projects high-dimensional data to 2D or 3D, preserving local neighbourhoods for visualisation. [Full guide →](/tools/t-sne)

**[TabICL](/glossary/tabicl)**: A tabular foundation model that adds column-then-row attention to scale in-context learning to bigger datasets. [Full guide →](/tools/tabicl)

**[TabNet](/glossary/tabnet)**: A deep model that selects features step by step with sequential attention. [Full guide →](/tools/tabnet)

**[TabPFN](/glossary/tabpfn)**: A transformer pre-trained on synthetic datasets that predicts in a single forward pass with in-context learning. [Full guide →](/tools/tabpfn)

**[TabSTAR](/glossary/tabstar)**: A tabular model that reads textual features through a pretrained language encoder together with numbers. [Full guide →](/tools/tabstar)

**[Tabular foundation models](/glossary/tabular-foundation-models)**: A large neural network pretrained across many tables so it can predict, impute or score new tables with little or no task-specific training. [Full guide →](/tools/tabular-foundation-models)

**[TD3](/glossary/td3)**: Improves DDPG with twin critics, delayed policy updates and target smoothing to curb value overestimation. [Full guide →](/tools/td3)

**[Telecom](/glossary/telecom)**: The domain of mobile and fixed networks, with huge event volumes and strict reliability needs. [Full guide →](/tools/telecom)

**[Temporal](/glossary/temporal)**: A platform that runs code as durable workflows that survive failures and retry automatically. [Full guide →](/tools/temporal)

**[Tensor parallelism](/glossary/tensor-parallelism)**: Divides individual weight matrices so several devices compute one layer together. [Full guide →](/tools/tensor-parallelism)

**[TensorFlow](/glossary/tensorflow)**: An end-to-end platform with a graph runtime and deployment tooling. [Full guide →](/tools/tensorflow)

**[TensorRT](/glossary/tensorrt)**: A compiler and runtime that optimizes neural networks for NVIDIA GPUs. [Full guide →](/tools/tensorrt)

**[Terraform](/glossary/terraform)**: A tool that declares cloud resources in configuration files and plans and applies changes. [Full guide →](/tools/terraform)

**[Test-time compute](/glossary/test-time-compute)**: Test-time compute is the extra computation a model spends while answering a single request, used to raise the quality of that answer. [Full guide →](/tools/test-time-compute)

**[Text-to-SQL](/glossary/text-to-sql)**: Using a language model to translate a natural-language question into an SQL query against a known schema. [Full guide →](/tools/text-to-sql)

**[TFT](/glossary/tft)**: An attention-based forecasting model with variable selection and interpretable temporal patterns. [Full guide →](/tools/tft)

**[Thales](/glossary/thales)**: Provider of key management, hardware security modules and tokenization products. [Full guide →](/tools/thales)

**[Tokenizer training](/glossary/tokenizer-training)**: Fitting a vocabulary from data so that inputs are represented efficiently. [Full guide →](/tools/tokenizer-training)

**[Torchtune](/glossary/torchtune)**: A PyTorch library of recipes for fine-tuning and evaluating language models with minimal abstraction. [Full guide →](/tools/torchtune)

**[Tracing](/glossary/tracing)**: Captures prompts, tool calls, inputs and outputs of every step as a trace. [Full guide →](/tools/tracing)

**[Transfer learning](/glossary/transfer-learning)**: Starts from weights learned on a large source task and adapts them to a smaller target task. [Full guide →](/tools/transfer-learning)

**[Transformers](/glossary/transformers)**: Neural networks built on self-attention, letting every element of the input weigh every other element. [Full guide →](/tools/transformers)

**[Tree-of-thoughts](/glossary/tree-of-thoughts)**: A search over multiple partial reasoning paths with evaluation and backtracking. [Full guide →](/tools/tree-of-thoughts)

**[Trino](/glossary/trino)**: A fast engine that queries data across lakes, databases and warehouses with standard SQL. [Full guide →](/tools/trino)

**[Triton](/glossary/triton)**: A server that runs models from many frameworks with dynamic batching and concurrent execution on GPUs. [Full guide →](/tools/triton)

**[TRL](/glossary/trl)**: Provides trainers for SFT, reward modelling, PPO, DPO and related alignment methods. [Full guide →](/tools/trl)

**[TRPO](/glossary/trpo)**: Constrains each policy update with a KL-divergence trust region, giving monotonic-improvement guarantees in theory. [Full guide →](/tools/trpo)

**[TruLens](/glossary/trulens)**: A library that records app runs and scores them with feedback functions for groundedness and relevance. [Full guide →](/tools/trulens)

**[TurboQuant](/glossary/turboquant)**: A quantization method from Google Research, published at ICLR 2026, that combines PolarQuant with a QJL residual step to compress vectors with little error. [Full guide →](/tools/turboquant)

## U

**[U-Net](/glossary/u-net)**: A convolutional encoder-decoder with skip connections that preserves fine detail for pixel-level predictions. [Full guide →](/tools/u-net)

**[UiPath](/glossary/uipath)**: A platform for building software robots that automate repetitive work in applications. [Full guide →](/tools/uipath)

**[UMAP](/glossary/umap)**: A manifold-learning method that projects high-dimensional data into two or three dimensions while preserving structure. [Full guide →](/tools/umap)

**[Unsloth](/glossary/unsloth)**: A library with optimized kernels that speeds up LoRA and QLoRA fine-tuning and cuts memory use. [Full guide →](/tools/unsloth)

**[Unsupervised learning](/glossary/unsupervised-learning)**: Learning from data without target labels, by discovering clusters, directions of variation, or density. [Full guide →](/tools/unsupervised-learning)

**[Urban intelligence](/glossary/urban-intelligence)**: Combines geospatial, property and regulatory data with AI to explain how land and buildings can be used. [Full guide →](/tools/urban-intelligence)

## V

**[VAE](/glossary/vae)**: An autoencoder that learns a distribution over a latent space and is trained with a variational objective. [Full guide →](/tools/vae)

**[Variational inference](/glossary/variational-inference)**: Turns Bayesian inference into optimization by fitting a simple distribution to a complex posterior. [Full guide →](/tools/variational-inference)

**[Vertex AI](/glossary/vertex-ai)**: A managed platform for training, tuning and deploying models and agents. [Full guide →](/tools/vertex-ai)

**[Vespa](/glossary/vespa)**: An engine that combines text, vector and structured search with on-node ranking. [Full guide →](/tools/vespa)

**[Vision-language-action models](/glossary/vision-language-action-models)**: A neural network that takes camera input and a natural-language instruction and outputs low-level actions for a robot. [Full guide →](/tools/vision-language-action-models)

**[ViT](/glossary/vit)**: Splits an image into patches and processes them as a sequence with a transformer. [Full guide →](/tools/vit)

**[vLLM](/glossary/vllm)**: An engine that uses PagedAttention and continuous batching to serve language models efficiently. [Full guide →](/tools/vllm)

**[VQ-VAE](/glossary/vq-vae)**: A variational autoencoder that quantizes its latent space into a learned codebook of discrete tokens. [Full guide →](/tools/vq-vae)

## W

**[Weaviate](/glossary/weaviate)**: A vector database that stores objects with embeddings and supports hybrid and filtered search. [Full guide →](/tools/weaviate)

**[Weight initialization](/glossary/weight-initialization)**: Weight initialization is the choice of starting values for a neural network's weights and biases before training begins. [Full guide →](/tools/weight-initialization)

**[Weights & Biases](/glossary/weights-and-biases)**: A platform that logs metrics, artifacts and sweeps for machine-learning runs. [Full guide →](/tools/weights-and-biases)

**[Whisper](/glossary/whisper)**: OpenAI's multilingual speech-to-text model trained on large-scale weak supervision. [Full guide →](/tools/whisper)

**[Word embeddings](/glossary/word-embeddings)**: Learned vector representations of words where similar words have similar vectors, as in Word2Vec and GloVe. [Full guide →](/tools/word-embeddings)

**[World models](/glossary/world-models)**: A learned model that predicts how an environment will change, given its current state and an action, so an agent can plan or train inside it. [Full guide →](/tools/world-models)

## X

**[XGBoost](/glossary/xgboost)**: A fast, regularized gradient-boosting library for tabular data. [Full guide →](/tools/xgboost)

**[XTab](/glossary/xtab)**: A framework that pre-trains shared transformer layers across many tables with table-specific embeddings. [Full guide →](/tools/xtab)

## Y

**[YOLO](/glossary/yolo)**: A family of single-pass detectors that predict boxes and classes quickly. [Full guide →](/tools/yolo)

## Z

**[Zero Trust](/glossary/zero-trust)**: A security model that authenticates and authorizes every request, regardless of network location. [Full guide →](/tools/zero-trust)

**[Zero-shot learning](/glossary/zero-shot-learning)**: The model solves a task it was never explicitly trained on, relying on general knowledge and a task description. [Full guide →](/tools/zero-shot-learning)
