THE AI ATLAS REFERENCE LIBRARY
A little clarity.
From A to Z.
A book-style index to the language of AI. Find a term, follow its link, and explore a plain-English explanation with a practical example.
145 terms · A–Z
A
B
C
D
E
F
G
H
I
J
K
L
M
N
O
P
Q
R
S
T
U
V
W
X
Y
Z
No terms match this search. Try an acronym, its full name, or Clear.
The numbers identify entries below, not page numbers. This is a growing reference, not an exhaustive list of every AI term. Select a term to read its explanation.
THE READING COMPANION
One term. A clearer understanding.
Open any entry, follow a related concept, or return to its place in the index. Examples are illustrative. Further-reading links lead to technical documentation and research.
001AccuracyEvaluation
Share of evaluated predictions that are correct; on imbalanced data it can hide poor performance on rare cases.
A classifier gets 90 of 100 labels right: 90% accuracy.
002Activation functionLearning
A rule applied inside a neural network that lets layers represent relationships beyond a simple linear mapping.
ReLU replaces negative inputs with zero.
003Adversarial exampleSafety
An input deliberately changed to make a model produce a wrong or misleading result.
A modified image causes an object classifier to mislabel it.
004AgentAgents
Also known as: AI agent
An AI system that can choose actions and use tools toward a goal, usually within permissions and stopping rules.
An assistant searches approved documents and drafts a report for review.
005Agentic AIAgents
An umbrella term for systems that pursue goals through actions, tool use and feedback; the degree of independence varies.
A research assistant decides which source to check next.
006AGIFoundations
Also known as: Artificial general intelligence
Artificial general intelligence: a proposed broad capability across many intellectual tasks. There is no universally accepted definition or test.
A high benchmark score alone does not establish AGI.
007AI governanceSafety
The policies, responsibilities and processes used to oversee how AI is developed, deployed and monitored.
An organisation assigns an owner, review process and incident procedure.
008AI literacyFoundations
The ability to understand AI capabilities and limits, use it appropriately and assess its outputs.
A learner checks a generated claim before sharing it.
009AlignmentSafety
Efforts to make an AI system behave consistently with intended goals, values and constraints; success is context-dependent.
Training an assistant to refuse requests outside its permitted task.
010APIDeployment
Also known as: Application programming interface
Application programming interface: a defined way for software systems to request functions or exchange data.
A website sends a text request to a model service.
011Artificial intelligenceFoundations
Also known as: AI
A broad field concerned with systems that perform tasks involving capabilities such as perception, prediction, language and planning.
Software identifies objects in an image.
012ASIFoundations
Also known as: Artificial superintelligence
Artificial superintelligence: a hypothetical capability substantially beyond human performance across a broad range of tasks.
A future scenario discussed in AI research, not a standard product certification.
013ASRLanguage & media
Also known as: Automatic speech recognition · Speech to text
Automatic speech recognition: converting spoken audio into written text.
Transcribing a recorded meeting into sentences.
014AttentionModels
A mechanism that weights information from different positions or inputs when building a representation.
A model connects a pronoun with a relevant earlier noun.
015AutoencoderModels
A neural network trained to reconstruct its input through an internal representation, often with a compact bottleneck.
Learning a compact representation of an image.
016Autoregressive modelModels
A model that generates a sequence by predicting each next element using earlier elements.
A language model continues a sentence token by token.
017BackpropagationLearning
A method for calculating how a network’s error changes with its parameters, allowing an optimiser to update them.
Finding which weights contributed to a prediction error.
018Batch sizeLearning
The number of examples processed together before a training update, or grouped for an inference operation.
Training updates use groups of 32 examples.
019BenchmarkEvaluation
A defined task set and evaluation procedure used to compare systems under specified conditions.
Comparing models on the same held-out question set.
020BiasSafety
A systematic tendency in data, modelling or outcomes; it can include harmful disparities between people or groups.
A hiring model performs differently across demographic groups.
021BPELanguage & media
Also known as: Byte pair encoding
Byte pair encoding: a subword tokenisation approach that repeatedly merges frequent adjacent symbols or token pairs.
A rare word is represented by several familiar pieces.
022Chain of thoughtPrompting
Intermediate reasoning steps expressed or generated during a task. Such text is not a guaranteed account of a model’s internal process.
A worked solution shows intermediate calculations for checking.
023ChatbotLanguage & media
Software that interacts through conversation; it may use rules, retrieval, language models or a combination.
A support assistant answers a question about opening hours.
024ChunkingRetrieval
Dividing a larger document into smaller pieces for processing or retrieval; boundaries and overlap affect results.
Splitting a manual into sections before indexing it.
025ClassificationLearning
Assigning an input to one or more predefined categories.
Labelling an email as spam or legitimate.
026ClusteringLearning
Grouping examples by similarity without requiring a predefined label for each group.
Finding clusters of similar customer questions.
027CNNModels
Also known as: Convolutional neural network
Convolutional neural network: a network using learned filters over local regions, commonly used for image or signal tasks.
Detecting visual patterns in a photograph.
028Computer visionLanguage & media
AI methods for interpreting visual information such as images and video.
Locating pedestrians in a road scene.
029Confusion matrixEvaluation
A table counting predicted classes against actual classes, making different kinds of mistakes visible.
Seeing how often a model confuses cats with dogs.
030Context engineeringPrompting
Designing the information available to a model, including instructions, retrieved material, conversation state and tool results.
Providing the relevant policy section instead of an entire document archive.
031Context windowLanguage & media
Also known as: Context length
The maximum sequence length a model can handle in one interaction; providers may count input, output and reasoning differently.
A long conversation leaves less room for a new document.
032Cosine similarityRetrieval
A measure of the angle between vectors, often used to compare embeddings; similarity does not establish factual agreement.
Ranking document vectors close to a question vector.
033Cross-validationEvaluation
Estimating performance by repeatedly training and evaluating on different partitions of the available data.
Rotating the held-out fold in five-fold validation.
034Data augmentationLearning
Creating additional training variations through controlled changes to existing examples.
Rotating training images to vary their orientation.
035Data driftDeployment
A change in the distribution of incoming data relative to a previous reference period.
Customer questions shift after a new product launch.
036Data leakageEvaluation
Information from evaluation data, future events or otherwise unavailable sources improperly enters training or decision-making.
A model sees test answers during training.
037DatasetFoundations
A collection of examples and associated information used for training, evaluation or analysis.
A set of images paired with object labels.
038DecoderModels
A component that produces or reconstructs an output from representations and context.
Generating text from previously processed information.
039Deep learningLearning
Machine learning using neural networks with multiple layers of learned representations.
A multilayer network recognises speech patterns.
040Diffusion modelModels
A generative model commonly trained to reverse a noise-adding process and produce structured samples.
Generating an image through a sequence of denoising steps.
041DistillationLearning
Training one model to learn from another model’s outputs or behaviour; a smaller student may trade capability for efficiency.
Training a compact classifier using a larger model’s predictions.
042DPOLearning
Also known as: Direct preference optimization · Direct preference optimisation
Direct preference optimisation: a training method that uses preferred and rejected responses to adjust a model’s behaviour.
Learning from pairs where reviewers preferred the clearer response.
043EmbeddingRetrieval
A numerical vector representing information in a form useful for comparison or learning; it is not a perfect summary of meaning.
Representing a question as numbers for document search.
044EncoderModels
A component that turns an input into a representation used by another part of a system.
Converting an image into features for a downstream classifier.
045EpochLearning
One full pass through a training dataset; training may use several passes or sample streams differently.
Processing each training example once completes an epoch.
046EvaluationEvaluation
Measuring a system against defined criteria using tests, human review or both; useful criteria depend on the task.
Checking answer correctness, source support and response time.
047ExplainabilitySafety
Methods for helping people understand factors behind a model’s output; an explanation can be incomplete or misleading.
Showing which input features influenced a prediction.
048F1 scoreEvaluation
The harmonic mean of precision and recall, useful when both kinds of classification error matter.
Comparing spam detectors using precision and recall together.
049FairnessSafety
Assessing and addressing unjustified disparities in AI outcomes. Different fairness criteria can conflict.
Checking whether error rates differ across relevant groups.
050FeatureLearning
An input variable or learned representation used by a model.
House size is a feature for estimating a house price.
051Feature engineeringLearning
Designing, transforming or selecting input features to make relevant patterns easier for a model to use.
Deriving day of week from a timestamp.
052Few-shot promptingPrompting
Providing a small number of examples in a prompt to demonstrate the desired task or format.
Showing two labelled examples before asking for a third label.
053Fine-tuningLearning
Further training an existing model on selected data to change its behaviour or adapt it to a task.
Adapting a model to a company’s preferred answer format.
054Foundation modelModels
A broadly trained model that can be adapted or used for multiple downstream tasks.
One language model supports summarisation and classification.
055GANModels
Generative adversarial network: a generator and discriminator are trained in competition to improve generated samples.
A generator produces images while a discriminator distinguishes generated from real ones.
056Generative AIFoundations
AI that produces content such as text, images, audio or code based on learned patterns and supplied inputs.
Producing a draft illustration from a description.
057Gradient descentLearning
An optimisation approach that adjusts parameters in a direction intended to reduce a loss.
Updating network weights after calculating a training error.
058Ground truthEvaluation
Reference labels or answers used for evaluation or training; their reliability depends on how they were established.
Expert-reviewed labels serve as a test reference.
059GroundingRetrieval
Connecting a response to supplied evidence or other reliable information; retrieved material still needs assessment.
Answering a question using a cited paragraph from a manual.
060GuardrailsSafety
Checks and controls around an AI system that constrain inputs, actions or outputs; they do not eliminate all risk.
Requiring approval before an assistant sends an email.
061HallucinationSafety
Generated content that is unsupported, misleading or incorrect, even when expressed confidently.
A model invents a journal article that does not exist.
062Human in the loopSafety
Also known as: HITL
A design in which people review, guide or approve particular parts of an automated process.
A person approves a draft before publication.
063HyperparameterLearning
A setting controlling a model or training procedure rather than a value learned in the ordinary training updates.
Choosing batch size or learning rate before a training run.
064In-context learningPrompting
Adapting a response to instructions or examples in the current context without updating model weights during that interaction.
Following a new classification pattern shown in a prompt.
065InferenceDeployment
Also known as: Model serving
Using a trained model to produce an output from new input; it is different from updating its training weights.
A deployed model classifies a newly uploaded image.
066Instruction tuningLearning
Fine-tuning on instruction-and-response examples to improve a model’s ability to follow requests.
Training on examples of questions paired with appropriate answers.
067JailbreakSafety
An attempt to bypass a model or application’s intended behavioural restrictions.
A prompt tries to persuade an assistant to ignore its safety rules.
068JSON schemaDeployment
A structured description of expected JSON data, including fields, types and constraints.
Requiring an output with a string title and an array of sources.
069Knowledge cutoffLanguage & media
A stated limit on training-data recency, which does not guarantee complete knowledge before that date or exclude later supplied context.
A model may need search to answer a question about a recent event.
070Knowledge graphRetrieval
A representation of entities and relationships, often useful for linking information across sources.
Connecting a company to its products and founders.
071KV cacheDeployment
Also known as: Key value cache
Stored attention keys and values reused during generation to avoid recomputing earlier sequence work.
A model reuses information from preceding tokens while continuing a response.
072LabelLearning
A target annotation associated with a training or evaluation example.
An image is annotated as containing a bicycle.
073Large language modelModels
Also known as: language model
A language model trained at large scale to represent and generate language; capabilities and limits vary across models.
Generating a draft summary of supplied text.
074LatencyDeployment
Time taken to produce a response or a stage of one, such as the first output token.
A chat application measures delay before the first visible word.
075Latent spaceModels
An internal representation space in which a model encodes learned factors or relationships.
An image generator works with a compressed representation before decoding pixels.
076Learning rateLearning
A setting that scales the size of parameter updates during training.
Reducing the update step when training becomes unstable.
077LLMModels
Also known as: Large language model
Abbreviation for large language model: a model trained at large scale on language-oriented data.
A text assistant uses an LLM to generate a response.
078LoRALearning
Also known as: Low-rank adaptation
Low-rank adaptation: a parameter-efficient fine-tuning method that learns smaller update matrices while typically keeping base weights frozen.
Adapting a model using a small set of additional trainable weights.
079Loss functionLearning
A rule that assigns a numerical penalty to predictions, providing a training objective to minimise.
Penalising predictions that assign low probability to the correct label.
080Machine learningFoundations
Methods that learn patterns from data to make predictions, decisions or representations rather than encoding every rule manually.
Learning to identify spam from labelled email examples.
081MCPAgents
Also known as: Model Context Protocol
Model Context Protocol: an open protocol for connecting AI applications with tools, resources and prompts. It does not itself make a connection trustworthy.
An application connects to an approved document-search server.
082Mixture of expertsModels
Also known as: MoE
An architecture that routes inputs to specialised model components; sparse variants activate only a subset of experts for each input.
A router selects expert networks to process a token.
083Model cardSafety
Documentation describing a model’s intended uses, evaluation, limitations and other relevant information.
A release includes a summary of tested use cases and known gaps.
084Model driftDeployment
Changes in a deployed model’s performance or behaviour over time, for example as the data or environment changes.
A fraud detector becomes less effective as fraud patterns change.
085MonitoringDeployment
Tracking a deployed system’s quality, errors, performance and other signals to detect problems.
Watching for rising failure rates and slower responses.
086Multi-agent systemAgents
A system with multiple agents assigned roles or interacting toward a task; more agents can add coordination cost and errors.
A researcher and reviewer exchange drafts under a controlled workflow.
087Multimodal AILanguage & media
AI that handles more than one modality, such as text, images, audio or video; supported combinations vary.
Asking a question about a supplied image.
088Neural networkModels
A model made from connected units whose parameters are adjusted to learn useful transformations of inputs.
Layers transform pixel values into an image label.
089NLPLanguage & media
Also known as: Natural language processing
Natural language processing: methods for analysing, understanding or generating human language.
Extracting named organisations from an article.
090Object detectionLanguage & media
Locating and classifying objects within an image or video frame.
Drawing boxes around cars in a street image.
091ObservabilityDeployment
The ability to understand a system’s behaviour using evidence such as logs, metrics and traces.
Following an agent’s tool calls to diagnose a failed task.
092OCRLanguage & media
Also known as: Optical character recognition
Optical character recognition: extracting machine-readable text from images or scanned documents.
Reading printed text from a scanned invoice.
093OntologyRetrieval
An explicit description of concepts, categories and relationships within a domain.
Defining how products, suppliers and orders relate.
094Open weightsModels
A release that makes trained model parameters available under stated licence terms; it does not necessarily include training data or unrestricted use.
Downloading permitted model weights for local inference.
095OrchestrationAgents
Coordinating model calls, tools, data and workflow steps, including retries, state and approval points.
A workflow retrieves evidence, drafts an answer and requests review.
096OverfittingLearning
Learning training-specific patterns or noise so well that performance on unseen data suffers.
A model memorises training examples but misses new ones.
097ParameterModels
A numerical value within a model that is typically learned during training, such as a network weight.
Training adjusts the weights that connect network units.
098PEFTLearning
Also known as: Parameter efficient fine tuning
Parameter-efficient fine-tuning: methods that adapt a model by training a relatively small subset of parameters or added components.
Using adapters instead of updating every base weight.
099Pre-trainingLearning
Initial broad training used to build a model before task-specific adaptation or post-training.
Learning language patterns from a large text corpus.
100PrecisionEvaluation
For a class, the proportion of predicted positives that are actually positive.
Eight of ten flagged spam emails are spam: 80% precision.
101PromptPrompting
Input supplied to a model to guide its response, such as a request, instructions, examples or context.
Summarise this paragraph in three bullet points.
102Prompt engineeringPrompting
Designing and refining prompts to improve task performance, then testing whether the changes help.
Specifying the task, constraints and required answer format.
103Prompt injectionSafety
An attempt to make an AI application follow attacker-controlled instructions, often embedded in material it processes.
A retrieved webpage tells an assistant to reveal private information.
104QuantisationDeployment
Also known as: Quantization
Representing model values with reduced numerical precision to lower storage or computation needs, sometimes affecting accuracy.
Storing selected model weights with fewer bits.
105RAGRetrieval
Also known as: Retrieval augmented generation
Retrieval-augmented generation: retrieving relevant material and supplying it as context to a generative model. It does not guarantee correctness.
Searching a company manual before drafting a cited answer.
106Reasoning modelModels
A model designed or trained to spend additional computation on multi-step tasks; extra reasoning effort does not guarantee a correct result.
Working through a difficult mathematical problem before giving an answer.
107RecallEvaluation
For a class, the proportion of actual positives that the system identifies.
Finding nine of ten spam emails gives 90% recall.
108Red teamingSafety
Structured adversarial testing intended to uncover vulnerabilities, unsafe behaviours or other failures.
Testing whether malicious retrieved text can trigger unauthorised tool use.
109RegressionLearning
Predicting a numerical quantity rather than a discrete class.
Estimating delivery time in minutes.
110RegularisationLearning
Also known as: Regularization
Methods that discourage unwanted complexity or improve generalisation during training.
Adding a weight penalty to reduce overfitting.
111Reinforcement learningLearning
Learning a policy through interactions and reward signals rather than only fixed input-label pairs.
An agent learns actions that improve a game score.
112ReLUModels
Rectified linear unit: an activation function that outputs zero for negative inputs and the input itself for positive inputs.
ReLU of −3 is 0; ReLU of 4 is 4.
113RerankingRetrieval
Reordering an initial set of retrieved candidates using an additional relevance assessment.
A second model ranks ten passages before three are used in an answer.
114Reward modelLearning
A model that estimates a preference or reward signal used to guide another learning or selection process.
Scoring candidate responses based on learned reviewer preferences.
115Risk assessmentSafety
Identifying possible harms, estimating their significance and choosing controls appropriate to the intended AI use.
Assessing the consequences of an incorrect automated decision.
116RLHFLearning
Also known as: Reinforcement learning from human feedback
Reinforcement learning from human feedback: using human preference information to guide reinforcement-learning updates to a model.
Reviewers compare responses and their preferences inform training.
117RobustnessSafety
The ability to maintain acceptable performance under relevant changes, noise, stress or adversarial conditions.
Testing an image classifier on blurry photographs.
118Self-attentionModels
Attention applied within a sequence so that positions can incorporate information from other positions, subject to the model’s mask.
A text representation uses relevant words from the same passage.
119Self-supervised learningLearning
Learning from targets constructed from the data itself rather than individually provided human labels.
Predicting a hidden or next token in a text sequence.
120Semantic searchRetrieval
Searching using representations of meaning rather than only exact keyword overlap; relevance still needs validation.
Finding reset instructions when a user asks how to regain access.
121Structured outputDeployment
An output constrained or checked against an expected format or schema; valid structure does not guarantee correct content.
Returning a JSON object with specified fields and types.
122Supervised learningLearning
Learning from examples paired with target labels or outputs.
Training a classifier with emails labelled spam or not spam.
123Synthetic dataLearning
Data generated or simulated for a purpose, often to augment training or testing; quality and privacy need assessment.
Creating simulated support questions for evaluating a search tool.
124TemperaturePrompting
A sampling setting that changes the shape of the output probability distribution; its effects depend on the system.
Comparing less variable and more varied candidate drafts.
125TensorModels
A numerical array with one or more dimensions used to represent data and model computations.
A batch of colour images is represented as a multidimensional array.
126Test setEvaluation
Data held apart from training and model selection for estimating final performance.
Measuring a finished classifier on examples it has not trained on.
127Text to speechLanguage & media
Generating spoken audio from written text.
Reading a navigation instruction aloud.
128ThroughputDeployment
The amount of work processed per unit of time, such as tokens or requests per second.
Measuring how many requests a service can handle each minute.
129TokenLanguage & media
A unit a model processes, which may be part of a word, punctuation, bytes or another modality-dependent representation.
A single word may be split into several text tokens.
130TokenisationLanguage & media
Also known as: Tokenization
Converting an input into a sequence of tokens according to a model’s encoding rules.
Splitting text into subword units before inference.
131Tool callingAgents
A model or application producing a structured request to invoke an external function; the application controls execution and permissions.
An assistant requests a weather lookup through an approved function.
132Top-pPrompting
Also known as: Nucleus sampling
Nucleus sampling: selecting from the smallest candidate set whose cumulative probability reaches a chosen threshold.
Limiting next-token candidates to a probability mass such as 0.9.
133TrainingLearning
Adjusting a model using data and an objective to learn parameters or behaviour.
Updating a network to reduce prediction error on examples.
134Training dataLearning
Examples used in the learning process; their quality, provenance and coverage influence model behaviour.
A collection of permitted documents used for language-model training.
135Transfer learningLearning
Reusing knowledge represented by a trained model for another related task or setting.
Adapting a pretrained vision model to identify plant diseases.
136TransformerModels
A neural-network architecture built around attention and other components; variants support language, vision and other tasks.
An attention-based model processes a sequence of text tokens.
137UnderfittingLearning
Failing to learn important patterns, often resulting in poor performance even on training data.
A model that is too simple misses a curved relationship.
138Unsupervised learningLearning
Learning structure from data without task-specific target labels for each example.
Grouping similar documents into clusters.
139Validation setEvaluation
Data used during development to choose settings or compare candidate models, separate from a final test set.
Choosing a learning rate based on validation performance.
140Vector databaseRetrieval
A system that stores vectors and supports similarity-based retrieval, often alongside metadata filters.
Searching document embeddings for passages related to a question.
141Vision-language modelLanguage & media
A model that connects visual and language information, with supported input and output forms depending on the model.
Answering a text question about a supplied photograph.
142WorkflowAgents
A defined arrangement of steps, conditions and tool calls for completing a task; it can include AI components.
Retrieve a document, draft a summary, then request human approval.
143XAISafety
Also known as: Explainable AI
Explainable artificial intelligence: approaches that help people examine and understand AI outputs and behaviour.
Inspecting feature contributions to a model prediction.
144YOLOLanguage & media
Also known as: You Only Look Once
You Only Look Once: a family of object-detection models designed to predict object locations and classes efficiently.
Identifying several objects within a camera frame.
145Zero-shot promptingPrompting
Asking a model to perform a task without demonstrating it through example input-output pairs in the prompt.
Classify this review as positive or negative, with no sample reviews provided.