Glossary
The words of artificial intelligence, explained plainly with definitions, practical examples, formulas and references for study, teaching and professional work.
450 termsContemporary models, video and systems
A
- MetricAccuracyProportion of correct predictions out of total predictions made.
- ConceptActivation FunctionA mathematical function applied to a neuron's output to introduce non-linearity into the network.
- ArchitectureActivation LayerA layer that applies a non-linear activation function element-wise to its input.
- AlgorithmAdagradAdaptive learning rate optimizer that adapts rates per parameter based on historical gradients.
- AlgorithmAdam OptimizerAn adaptive learning rate optimization algorithm combining momentum and RMSprop.
- OptimisationAdamWAn Adam variant separating weight decay from the adaptive update based on gradient moments.
- Evaluation & safetyAdversarial exampleAn input intentionally modified to induce model error, often with small perturbations under a chosen metric.
- Evaluation & safetyAdversarial promptAn input designed to cause errors, evasion or undesired behavior in a model.
- TechniqueAdversarial TrainingTraining technique improving model robustness by including adversarial examples.
- Agenti e protocolliAgent cardA machine-readable document describing an agent’s identity, capabilities, endpoints and interaction methods.
- AgentsAgent memoryInformation retained across steps or sessions to preserve state, preferences and useful results for an agent.
- AgentsAgent orchestrationCoordination of state, tools, messages and stopping criteria across one or more agents.
- AgentsAgent workflowA controlled sequence of model and tool operations, with autonomy constrained by explicit rules.
- AgentsAgent2Agent (A2A)A protocol for exchanging tasks, messages and artifacts between agents built by different vendors.
- RAG & retrievalAgentic RAGRAG in which an agent plans searches, tools and checks instead of following a single fixed pipeline.
- AgentsAI agentA system that uses a model and tools to pursue a goal through successive decisions and actions.
- Video e media generativiAI avatarAn animated visual representation of a person or character used for interaction, presentation or automated assistance.
- InferenceAI gatewayAn infrastructure layer that centralizes access, routing, security and observability for AI services.
- Evaluation & safetyAI governanceThe roles, processes and controls through which an organization manages AI responsibilities and risks.
- FoundationsAleatoric uncertaintyUncertainty caused by intrinsic noise or variability in data, which cannot be removed by collecting more.
- ArchitectureAlexNetDeep CNN that won ImageNet 2012, pioneering deep learning in computer vision.
- ConceptAlgorithmA step-by-step procedure or formula for solving a problem or performing a task.
- LLMs & promptingAlignmentThe process of making an AI system’s behavior and objectives consistent with specified intentions, values and constraints.
- Hardware e infrastrutturaApple Neural EngineAn accelerator integrated into Apple chips for efficient inference on models compatible with platform software and formats.
- ConceptArtificial Intelligence (AI)Field of computer science focused on creating systems capable of performing tasks requiring human intelligence.
- Generative AIASR (Automatic Speech Recognition)Automatically converting speech into text through acoustic and language models.
- RepresentationAttention maskA matrix or vector limiting which positions may contribute to attention for each token.
- TechniqueAttention MechanismA technique allowing models to focus on specific parts of the input when producing output.
- ConceptAttention ScoreComputed similarity between query and key vectors before softmax normalization in attention.
- ConceptAttention WeightA scalar value indicating how much focus to place on a specific part of the input when producing output.
- MetricAUC-ROCArea under ROC curve, measuring binary classifier quality across all thresholds.
- ArchitectureAutoencoderA neural network trained to reconstruct its input, learning compressed representations in the process.
- OptimisationAutomatic differentiationSystematic calculation of exact program derivatives by applying rules to elementary operations in the computational graph.
- TechniqueAverage PoolingA pooling operation that computes the average value from each window of the feature map.
B
- Evaluation & safetyBackdoor attackA compromise inducing attacker-chosen behaviour when a specific trigger appears in the input.
- TrainingBackpropagationAn algorithm for training neural networks by calculating gradients of the loss function with respect to weights.
- Evaluation & safetyBalanced accuracyThe average of recall computed separately for every class, useful when class supports are highly imbalanced.
- AlgorithmBatch Gradient DescentA gradient descent variant that computes gradients using the entire training dataset in each iteration.
- TechniqueBatch NormalizationNormalizes layer inputs using batch statistics to stabilize and accelerate training.
- ConceptBatch SizeNumber of training examples processed together in one forward/backward pass.
- ProbabilityBayes’ theoremA rule relating conditional probabilities by reversing conditioning through likelihood, prior probability and evidence.
- AlgorithmBeam SearchDecoding algorithm keeping top B most likely sequences at each step.
- Evaluation & safetyBenchmark contaminationThe presence in training data of examples identical or similar to those used for model evaluation.
- ProbabilityBernoulli distributionA probability model for a binary variable taking value one with probability p and zero with probability one minus p.
- RAG & retrievalBi-encoderAn architecture that encodes queries and documents separately into vectors comparable by similarity.
- ConceptBias TermAn additional learnable parameter in neural networks that allows shifting the activation function.
- MetricBLEU ScoreMetric for machine translation quality comparing n-gram overlap between generated and reference translations.
- RAG & retrievalBM25A lexical ranking function that scores documents using term frequency, rarity and document length.
- ProbabilityBootstrapA resampling-with-replacement method used to estimate variability and intervals without a simple analytical formula.
- Evaluation & safetyBrier scoreThe mean squared difference between predicted probabilities and observed outcomes used to assess probabilistic accuracy.
- Agenti e protocolliBrowser agentAn agent that interprets web pages and interacts with navigation, forms and content to complete online operations.
C
- FoundationsCalibrationThe degree to which predicted confidence matches the frequency with which predictions are correct.
- LearningCatastrophic forgettingLoss of performance on previous knowledge when a model is updated on new data or tasks.
- ProbabilityCategorical distributionA discrete distribution assigning probability to each of finitely many mutually exclusive categories.
- LearningCausal language modellingAn objective training a model to predict the next token using exclusively the preceding context.
- TechniqueCausal MaskingMasking technique preventing attention to future positions in autoregressive models.
- ProbabilityCentral limit theoremA result showing that normalised sums or averages of many independent variables approach a normal distribution under suitable conditions.
- Hardware e infrastrutturaCerebras WSEA Cerebras wafer-scale processor integrating a very large number of cores and memory on a single wafer.
- OptimisationChain ruleThe differentiation rule for composite functions, multiplying local sensitivities along the computation path.
- TechniqueChain-of-Thought PromptingPrompting technique encouraging LLMs to show intermediate reasoning steps before answering.
- RAG & retrievalChunkingSplitting documents into smaller segments for indexing, retrieval and insertion into model context.
- FoundationsClass imbalanceA highly unequal class distribution that can make accuracy and training misleading.
- TaskClassificationA supervised learning task where the goal is to predict discrete class labels for input data.
- Generative AIClassifier-free guidanceGuidance that combines conditional and unconditional predictions to increase prompt adherence.
- Modelli e prodotti AIClaudeAn Anthropic model family used for text, code, multimodal analysis and agentic applications.
- ModelCLIP (Contrastive Language-Image Pre-training)A multimodal model trained to understand relationships between images and text.
- TaskClusteringAn unsupervised learning task that groups similar data points together based on their features.
- ArchitectureCNN (Convolutional Neural Network)A deep learning architecture specialized for processing grid-like data such as images, using convolutional layers.
- DatasetCOCO (Common Objects in Context)Large-scale dataset for object detection, segmentation, and captioning with 330k images.
- AgentsCode agentAn agent specialized in analyzing, modifying and verifying code using repositories and development tools.
- Agenti e protocolliCoding agentAn agent that explores repositories, changes code, runs commands and checks results using development tools.
- Evaluation & safetyCohen’s kappaAn agreement metric correcting observed concordance for agreement expected by chance from marginal distributions.
- Modelli e prodotti AICommand RA Cohere model series designed for enterprise applications, retrieval augmented generation and tool use.
- OptimisationComputational graphA representation of computation as operations and dependencies connecting inputs, parameters, intermediate values and outputs.
- AgentsComputer useAn agent’s ability to operate graphical interfaces through observation and actions such as clicking and typing.
- FieldComputer VisionA field of AI enabling computers to derive meaningful information from visual inputs like images and videos.
- FoundationsConcept driftA change over time in the relationship between inputs and expected outcomes that ages learned behavior.
- ProbabilityConditional probabilityThe probability of an event given another event, written P(A|B) and defined by the ratio between joint probability and P(B).
- ProbabilityConfidence intervalAn interval produced by a procedure that, repeated across samples, contains the parameter at a stated nominal frequency.
- ProbabilityConformal predictionA framework constructing prediction sets or intervals with coverage guarantees under exchangeability assumptions.
- ConceptConfusion MatrixA table used to evaluate classification model performance by showing true vs predicted classes.
- LLMs & promptingConstitutional AIAn alignment approach that uses an explicit set of principles to guide response revision and learning.
- AI contemporaneaContext engineeringDesigning the information, instructions, memory and tools supplied to a model at each stage of a workflow.
- LLMs & promptingContext windowThe maximum number of tokens a model can consider at once across instructions, context and output.
- InferenceContinuous batchingScheduling that adds and removes requests during generation instead of waiting for the whole batch.
- ParadigmContrastive LearningSelf-supervised learning contrasting positive pairs against negative pairs.
- Generative AIControlNetAn architecture that adds spatial conditions to a diffusion model without replacing its core weights.
- ConceptConvolutionA mathematical operation that slides a filter/kernel over input data to extract features.
- ArchitectureConvolutional LayerA layer in CNNs that applies convolution operations to extract spatial features from input data.
- InferenceCore MLAn Apple framework for running models on Apple devices using CPU, GPU and Neural Engine.
- RAG & retrievalCorrective RAGA RAG pipeline that evaluates retrieved material and corrects its course before generating an answer.
- RepresentationCorrelationA normalised measure of the relationship between variables, often bounded between minus one and one for Pearson’s coefficient.
- OptimisationCosine annealingA schedule reducing learning rate along a cosine curve from a maximum towards a predefined minimum.
- RepresentationCovarianceA joint measure of how two variables vary around their means, with sign related to relationship direction.
- RAG & retrievalCross-encoderA model that jointly processes query and document to estimate their relevance accurately.
- TechniqueCross-ValidationResampling technique evaluating model performance by splitting data into multiple train-test folds.
- FoundationsCurriculum learningA training strategy that presents examples in a designed order, often from easier to harder ones.
D
- Immagini generativeDALL·EAn OpenAI model series for creating and editing images from textual prompts and visual references.
- TechniqueData AugmentationTechniques to artificially increase training data size by creating modified versions of existing data.
- FoundationsData leakageThe improper use during training or evaluation of information unavailable at prediction time.
- OptimisationData parallelismA distributed strategy replicating the model and assigning different batches to devices before synchronising gradients.
- Evaluation & safetyData poisoningManipulation of training or continuous-learning data to intentionally alter a model.
- ConceptDatasetA collection of data examples used for training, validation, or testing machine learning models.
- Generative AIDDIM (Denoising Diffusion Implicit Model)A diffusion sampling method that enables non-Markovian trajectories and fewer steps.
- Generative AIDDPM (Denoising Diffusion Probabilistic Model)A diffusion model that learns to gradually reverse a noise-addition process.
- LearningDecision treeA model recursively partitioning feature space through rules until producing a prediction at the leaves.
- ProbabilityDeep ensembleA set of independently trained neural models whose predictions are aggregated to improve robustness and uncertainty.
- ConceptDeep LearningA subset of machine learning using neural networks with multiple layers to learn hierarchical representations of data.
- Agenti e protocolliDeep researchA workflow in which an agent searches, reads, compares and synthesises multiple sources to produce a documented report.
- Modelli e prodotti AIDeepSeekA model family from Chinese company DeepSeek, also known for reasoning variants and efficient architectures.
- Evaluation & safetyDefence in depthA strategy combining independent controls across layers so one failure does not compromise the whole system.
- Evaluation & safetyDemographic parityA criterion comparing positive-decision frequency across groups independently of actual class.
- RAG & retrievalDense retrievalDocument retrieval by comparing dense vector representations of queries and content.
- TechniqueDepthwise Separable ConvolutionDepthwise separable convolution factorizes a standard convolution into spatial filtering per channel followed by channel mixing. It is a common building block in efficient convolutional networks.
- TechniqueDice LossLoss function based on Dice coefficient, commonly used for segmentation tasks.
- Evaluation & safetyDifferential privacyA formal guarantee limiting how much one individual’s presence in data can influence the result.
- ArchitectureDiffusion ModelsGenerative models that learn to create data by reversing a gradual noising process.
- TechniqueDilated ConvolutionConvolution with gaps between kernel elements, increasing receptive field without adding parameters.
- TechniqueDimensionality ReductionTechniques to reduce the number of features in data while preserving important information.
- Evaluation & safetyDisparate impactA disproportionate difference in outcomes between groups that can arise without explicit discriminatory intent.
- FoundationsDistribution shiftA change in data distribution between training and deployment that can degrade performance.
- RepresentationDot productThe sum of products between corresponding vector components, related to angle, norms and projections.
- LLMs & promptingDPO (Direct Preference Optimization)A method that directly optimizes a policy on preferred and dispreferred response pairs without explicit RL.
- Generative AIDreamBoothA customization technique that specializes a diffusion model on a subject using a few images.
- TechniqueDropoutA regularization technique that randomly deactivates neurons during training to prevent overfitting.
E
- TechniqueEarly StoppingA regularization technique that stops training when validation performance stops improving.
- Hardware e infrastrutturaEdge AIInference execution on devices close to data, such as phones, sensors or industrial computers, rather than cloud only.
- Evaluation & safetyEffect sizeA measure of the magnitude of a difference or relationship, distinct from statistical significance and sample size.
- RepresentationEigenvalues and eigenvectorsScalars and directions for which a linear transformation changes only vector scale without altering direction.
- TechniqueEmbeddingAn embedding is a learned numerical vector that represents an item such as a word, sentence, image or product. Models place items in a shared space so that geometric relations can support comparison and retrieval.
- Generative AIEmbodied AIAI that perceives and acts through a physical or simulated body while interacting with an environment.
- ArchitectureEncoder-DecoderAn architecture where an encoder processes input into a representation and a decoder generates output from it.
- ProbabilityEntropyA measure of the average uncertainty of a probability distribution, maximal when outcomes are evenly distributed.
- FoundationsEpistemic uncertaintyUncertainty caused by limited model knowledge, potentially reducible with better data or information.
- ConceptEpochOne complete pass through the entire training dataset during training.
- Evaluation & safetyEqualized oddsA fairness criterion requiring comparable error rates across groups conditional on the actual class.
- ConceptErrorThe difference between predicted output and true label, indicating model's mistakes.
- RepresentationEuclidean distanceThe distance between points calculated as the square root of summed squared differences between corresponding coordinates.
- ProbabilityExpected valueThe probability-weighted average of a random variable, also called mathematical expectation.
- Evaluation & safetyExplainabilityThe ability to provide understandable explanations of factors and processes associated with a model result.
- ConceptExploding GradientA problem where gradients become extremely large during training, causing unstable updates and divergence.
F
- MetricF1 ScoreThe harmonic mean of precision and recall, providing a single balanced metric.
- Evaluation & safetyFairnessEvaluating and managing a model’s impacts and disparities on individuals and groups in a defined context.
- RAG & retrievalFAISSA library for vector similarity and clustering that implements efficient indexes for large collections.
- ConceptFeatureAn individual measurable property or characteristic of data used as input to a model.
- TechniqueFeature ExtractionThe process of transforming raw data into numerical features that machine learning models can process.
- ConceptFeature MapThe output of applying a convolution filter to an input, representing detected features.
- RepresentationFeature scalingTransforming feature scales to make distances, gradients and numerical contributions comparable during training or analysis.
- RepresentationFeature selectionThe process of selecting a feature subset to simplify models and reduce noise, cost or redundancy.
- ConceptFew-Shot LearningA model's ability to learn from a small number of examples, typically 1-10 examples per class.
- ConceptFilter / KernelA small matrix of learnable weights that slides over input during convolution to detect specific features.
- TrainingFine-tuningThe process of adapting a pre-trained model to a specific task by continuing training on task-specific data.
- OptimisationFlashAttentionAn attention algorithm reducing memory transfers through tiling without materialising the full score matrix.
- Immagini generativeFLUXA Black Forest Labs image generation and transformation model family with different variants and usage conditions.
- TechniqueFocal LossFocal loss is a modified cross-entropy objective for classification with severe class imbalance. It reduces the contribution of examples the model already classifies confidently and correctly.
- ConceptFoundation ModelLarge-scale models trained on broad data that can be adapted to a wide range of downstream tasks.
- Video e media generativiFrame interpolationGeneration of intermediate frames between existing images to increase smoothness or change a sequence’s frame rate.
- ArchitectureFully Connected LayerA neural network layer where every neuron is connected to every neuron in the previous and next layers.
- LLMs & promptingFunction callingA mechanism by which a model generates the name and structured arguments of an application-provided function.
G
- ArchitectureGAN (Generative Adversarial Network)A framework where two neural networks compete: a generator creates fake data and a discriminator tries to distinguish real from fake.
- Modelli e prodotti AIGeminiA Google DeepMind multimodal family for processing and generating content across text, images, audio and other channels.
- Modelli e prodotti AIGemmaA Google open-model family based on Gemini research and designed for developers and flexible deployments.
- InferenceGGUFA file format for distributing models and metadata, used primarily by the llama.cpp ecosystem.
- ModelGPT (Generative Pre-trained Transformer)A family of large language models developed by OpenAI that use transformer architecture for text generation.
- Modelli e prodotti AIGPT-5An OpenAI model family designed for reasoning, generation and tool use across different application contexts.
- ConceptGradientDirection and magnitude of steepest increase in loss function with respect to parameters.
- FoundationsGradient accumulationAccumulating gradients over several mini-batches before updating weights, emulating a larger batch.
- LearningGradient boostingAn ensemble method sequentially adding weak models to correct errors according to a loss gradient.
- OptimisationGradient checkpointingA technique storing selected activations and recomputing others during backpropagation to reduce training memory.
- AlgorithmGradient DescentAn optimization algorithm that iteratively adjusts parameters to minimize a loss function by following the gradient.
- Modelli e prodotti AIGraniteAn IBM model family oriented towards enterprise cases, code, documents and governed deployments.
- RAG & retrievalGraphRAGA RAG approach that integrates graphs of entities and relations to retrieve context and synthesize connections.
- Modelli e prodotti AIGrokAn xAI model family integrated into conversational products and also oriented towards reasoning and tools.
- Hardware e infrastrutturaGroq LPUA Groq architecture optimised for deterministic, low-latency inference with execution and memory organised differently from GPUs.
- LLMs & promptingGroundingConnecting model responses to verifiable sources, data or observations rather than linguistic plausibility alone.
- TechniqueGroup NormalizationNormalizes by dividing channels into groups and normalizing within each group.
- ArchitectureGRU (Gated Recurrent Unit)Simplified RNN variant with gating mechanisms, similar to LSTM but fewer parameters.
- LLMs & promptingGuardrailsTechnical and procedural controls that constrain an AI system’s inputs, outputs and actions to reduce risk.
H
- LLMs & promptingHallucinationPlausible content unsupported by facts or available sources, generated as if it were correct.
- OptimisationHessian matrixThe matrix of second derivatives of a scalar function, describing local curvature across variable pairs.
- ArchitectureHidden LayerLayers in a neural network between input and output that learn intermediate representations.
- RAG & retrievalHNSW (Hierarchical Navigable Small World)An approximate nearest-neighbor index that organizes vectors in a multilayer graph.
- OptimisationHuber lossA regression loss quadratic for small residuals and linear beyond a delta threshold to limit outlier influence.
- AgentsHuman-in-the-loopA design that inserts human review or decision-making at meaningful points in an automated process.
- Video e media generativiHunyuanVideoA Tencent video model with public implementations for studying cinematic generation and multimodal architectures.
- RAG & retrievalHybrid searchRetrieval that combines lexical and semantic rankings to exploit exact matches and meaning similarity.
- RAG & retrievalHyDE (Hypothetical Document Embeddings)A technique that generates a hypothetical answer and uses its embedding to retrieve real documents.
- ConceptHyperparameterA configuration variable that is set before training and controls the learning process.
- Evaluation & safetyHypothesis testingA statistical procedure comparing data with a null hypothesis to assess evidence against a reference model.
I
- Generative AIImage-to-imageTransforming an existing image guided by prompts, conditions or modification strength.
- Video e media generativiImage-to-videoGeneration of motion and new frames from one or more images used as visual references.
- Immagini generativeImagenA Google model family for image generation from textual descriptions and visual editing tasks.
- DatasetImageNetLarge-scale image dataset with 14M images across 20k categories, used for ILSVRC competition.
- ProbabilityImportance samplingA technique sampling from a convenient distribution and reweighting outcomes to estimate quantities under another distribution.
- ArchitectureInception ModuleCNN building block applying multiple filter sizes in parallel and concatenating results.
- ConceptInferenceThe process of using a trained model to make predictions on new data.
- Generative AIInpaintingGenerating or reconstructing selected areas of an image while preserving surrounding context.
- ConceptInputData fed into a model or neural network for processing.
- ArchitectureInput LayerThe first layer of a neural network that receives raw input data.
- TechniqueInstance NormalizationNormalizes each sample independently, commonly used in style transfer.
- Evaluation & safetyIntegrated gradientsA method assigning prediction attribution by integrating gradients along a path from a baseline.
- Evaluation & safetyInterpretabilityThe degree to which a model’s structure and behavior can be directly understood and analyzed.
- Immagini generativeIP-AdapterAn adapter conditioning a diffusion model with reference images to guide subject, style or composition.
- RAG & retrievalIVF (Inverted File Index)A vector index that partitions space into clusters and limits search to a subset.
J
K
- LearningK-nearest neighboursA method predicting from the k nearest training examples under a distance metric, without a traditional parametric fitting phase.
- RepresentationKernel trickA technique computing dot products in an implicit feature space without explicitly constructing every coordinate.
- Modelli e prodotti AIKimi K2A Moonshot AI model in the Kimi series designed for reasoning, coding and agentic activities.
- Video e media generativiKlingA Kuaishou model family for text- or image-driven video generation with visual and motion controls.
- LLMs & promptingKV cacheA temporary store of key and value tensors that avoids recomputing them for previous tokens during generation.
L
- ConceptLabelThe target output or ground truth associated with a training example in supervised learning.
- RepresentationLate interactionA retrieval architecture comparing multiple token representations after separate query and document encoding.
- Generative AILatent diffusionDiffusion performed in a compressed latent space rather than directly on pixels.
- LearningLearning curveA performance chart across data volume or iterations used to observe convergence, underfitting and overfitting.
- TrainingLearning RateA hyperparameter controlling how much model weights are updated during training.
- ConceptLearning RateHyperparameter controlling how much to adjust model weights during training.
- TechniqueLearning Rate ScheduleA strategy for adjusting the learning rate during training to improve convergence and performance.
- OptimisationLearning rate warmupAn initial phase in which learning rate gradually increases to the value intended by the subsequent schedule.
- Evaluation & safetyLeast privilegeThe principle granting users and systems only permissions and access strictly necessary for the intended task.
- ProbabilityLikelihoodA function of parameters measuring how probable a model makes the observed data while keeping observations fixed.
- Evaluation & safetyLIMEA technique that explains a prediction by locally approximating the model with an interpretable one.
- LearningLinear regressionA model estimating a continuous target as a linear combination of features plus an intercept term.
- Video e media generativiLip syncAlignment of a face’s mouth movements with the phonemes and timing of a vocal track.
- Modelli e prodotti AILlamaA Meta language-model family released in variants of different sizes and capabilities, including downloadable weights.
- Inferencellama.cppA runtime optimized for running language models locally on CPUs and supported accelerators.
- ModelLLM (Large Language Model)A neural network with billions of parameters trained on massive text datasets to understand and generate human language.
- Evaluation & safetyLLM evaluationSystematic measurement of a language model’s quality, safety, cost and behavior on defined tasks.
- Evaluation & safetyLLM-as-a-judgeUsing a model to evaluate other models’ responses according to an explicit rubric.
- Evaluation & safetyLog lossA loss penalising probability assigned away from the correct class through the negative logarithm of correct confidence.
- LearningLogistic regressionA linear classification model transforming a score through sigmoid or softmax to estimate class probabilities.
- TechniqueLoRA (Low-Rank Adaptation)Parameter-efficient fine-tuning adding trainable low-rank matrices to frozen weights.
- ConceptLossA measure of how wrong the model's predictions are, used to guide training.
- ConceptLoss FunctionA function that measures the difference between predicted and actual values, guiding model optimization.
- Video e media generativiLuma Dream MachineA Luma AI system for generating video and visual transformations from prompts and multimedia references.
M
- TaskMachine TranslationAutomatic translation of text from one language to another.
- MetricMAE (Mean Absolute Error)Loss function measuring average absolute difference between predicted and actual values.
- RepresentationManhattan distanceThe sum of absolute coordinate differences, also called L1 or city-block distance.
- ProbabilityMarkov chainA process in which the next-state distribution depends only on the present state rather than earlier history.
- TaskMasked Language ModelingMasked language modeling trains a model to reconstruct selected tokens from their surrounding context. It is the main pretraining objective used by BERT-style encoder models.
- ConceptMatrixA 2D array of numbers arranged in rows and columns.
- RepresentationMatrix rankThe maximum number of linearly independent rows or columns, equal to the dimension of a transformation’s image.
- TechniqueMax PoolingA pooling operation that takes the maximum value from each window of the feature map.
- LearningMaximum likelihoodAn estimation method choosing parameters that maximise the probability assigned by the model to observed data.
- Evaluation & safetyMean reciprocal rankThe mean reciprocal position of the first relevant result across a set of queries.
- Evaluation & safetyMembership inferenceAn attack that estimates whether a particular example was present in a model’s training data.
- Agenti e protocolliMemory consolidationThe process of summarising, merging and selecting agent memories to build more compact, useful persistent memory.
- RAG & retrievalMetadata filteringRestricting vector or lexical search using structured document attributes.
- Modelli e prodotti AIMistralA model family from French company Mistral AI for language, coding and commercial or self-managed deployments.
- FoundationsMixed precisionTraining or inference that combines numerical formats of different precision to reduce memory and compute.
- TechniqueMixupData augmentation creating synthetic examples by mixing pairs of training samples.
- InferenceMLXAn Apple numerical framework for machine learning with arrays and graphs optimized for Apple silicon.
- Hardware e infrastrutturaMLXAn Apple numerical-computing and machine-learning framework for Apple silicon with accelerated arrays and operations.
- ArchitectureMobileNetEfficient CNN architecture using depthwise separable convolutions for mobile deployment.
- ConceptModelA mathematical representation learned from data that makes predictions or decisions.
- Evaluation & safetyModel cardA document describing a model’s purpose, data, evaluation, limitations and conditions of use.
- FoundationsModel collapseThe progressive degradation of a generative model trained on synthetic data produced by previous generations.
- AgentsModel Context Protocol (MCP)An open protocol for connecting AI applications to tools, resources and prompts through a standard interface.
- Evaluation & safetyModel inversionAn attack that reconstructs features or examples from training data by exploiting model outputs.
- Evaluation & safetyModel monitoringContinuous observation of quality, data, drift and operational behaviour after model deployment.
- AI contemporaneaModel routerA component routing each request to a model according to task, cost, capability, latency or operating rules.
- InferenceModel routingDynamic selection of a model or execution path based on request, cost and constraints.
- TechniqueMomentumAn optimization technique that accelerates gradient descent by accumulating a velocity vector in directions of persistent reduction in the loss.
- ProbabilityMonte Carlo methodA family of methods estimating quantities through repeated random sampling and averages of the resulting outcomes.
- Video e media generativiMotion transferTransfer of motion from a source video or performance to a different visual subject while preserving its identity.
- MetricMSE (Mean Squared Error)Loss function measuring average squared difference between predicted and actual values.
- Agenti e protocolliMulti-agent debateA method in which multiple agents produce arguments and critiques before selecting or synthesising a final response.
- AgentsMulti-agent systemA system in which multiple agents with distinct roles or objectives collaborate or negotiate to complete a task.
- Evaluation & safetyMulticollinearityStrong linear dependence among predictors that destabilises coefficients and attributions in statistical models.
- ConceptMultimodal AIAI systems that can process and relate information from multiple modalities like text, images, audio, and video.
- Audio e multimodaleMusic generationSynthesis of melodies, arrangements or complete tracks by models learning rhythmic, harmonic and timbral structure.
N
- LearningNaive BayesA probabilistic classifier applying Bayes’ rule under conditional feature independence given the class.
- Audio e multimodaleNative multimodal modelA model designed and trained jointly across modalities rather than merely connecting separate components afterwards.
- TaskNER (Named Entity Recognition)NLP task identifying and classifying named entities (persons, organizations, locations) in text.
- ArchitectureNeural NetworkA computational model inspired by biological neural networks, consisting of interconnected nodes (neurons) that process information.
- Hardware e infrastrutturaNeural processing unitAn accelerator designed for artificial-intelligence operations with emphasis on energy efficiency and local inference.
- ConceptNeuronBasic computational unit in neural networks that receives inputs, applies weights and activation, produces output.
- TaskNext Sentence PredictionPre-training task predicting whether sentence B follows sentence A.
- FoundationsNoisy labelsIncorrect or inconsistent labels in data that can distort learning and evaluation.
- ProbabilityNormal distributionA symmetric continuous distribution determined by mean and variance, with the characteristic bell shape also called Gaussian.
- Evaluation & safetyNormalised discounted cumulative gainA ranking metric rewarding relevance and position with logarithmic discount, normalised against the ideal order.
- TechniqueNucleus Sampling (Top-P)Text generation sampling from smallest set of tokens whose cumulative probability exceeds threshold P.
O
- TaskObject DetectionA computer vision task that identifies and localizes objects within an image using bounding boxes.
- Hardware e infrastrutturaOllamaA tool for downloading, configuring and running local language models with a simplified CLI and API.
- Modelli e prodotti AIOLMoAn AI2 model family strongly emphasising openness of weights, data and the training process.
- RepresentationOne-hot encodingA representation of categories using binary vectors with only one active position for each observed value.
- InferenceONNXAn open format for representing models and transferring them between frameworks and inference runtimes.
- Modelli e prodotti AIOpen-weight modelA model whose weights are available for download and independent use, under conditions defined by the distributor’s licence.
- ConceptOptimizationThe process of adjusting model parameters to minimize the loss function and improve performance.
- RepresentationOrthogonalityThe relationship between vectors whose dot product is zero, geometrically interpreted as perpendicularity in Euclidean space.
- FoundationsOut-of-distribution (OOD)A condition where an input differs sufficiently from training data to make prediction unreliable.
- ConceptOutputThe result produced by a model after processing input data.
- ArchitectureOutput LayerThe final layer of a neural network that produces predictions or outputs.
- ConceptOverfittingWhen a model learns training data too well, including noise, resulting in poor generalization to new data.
P
- Evaluation & safetyP-valueUnder the null hypothesis, the probability of observing a result at least as extreme as the one obtained by the test.
- TechniquePaddingAdding extra pixels around the border of input data to control output size in convolution operations.
- RepresentationPadding maskA mask excluding padding tokens from operations so they do not affect attention, loss or aggregation.
- InferencePaged attentionManaging the KV cache in non-contiguous blocks to reduce fragmentation and memory waste.
- ConceptParameterLearnable values in a model that are optimized during training (weights and biases).
- RepresentationParent document retrievalA strategy retrieving small segments while returning the broader parent document or context to the model.
- ArchitecturePerceptronSimplest neural network with single layer, binary classifier invented in 1950s.
- MetricPerplexityMeasurement of how well a probability model predicts a sample, for evaluating language models.
- Modelli e prodotti AIPhiA Microsoft series of relatively compact models designed for language and multimodal capabilities with reduced requirements.
- Video e media generativiPikaA platform for creating and editing video with generative models, effects and text- or image-guided transformations.
- OptimisationPipeline parallelismDistribution of layer groups across devices with micro-batches flowing through the pipeline in sequence.
- AgentsPlanningDecomposing a goal into a sequence or structure of executable and verifiable actions.
- ArchitecturePooling LayerA downsampling layer in CNNs that reduces spatial dimensions while retaining important features.
- TechniquePositional EncodingA technique to inject position information into transformer inputs since transformers lack inherent sequence order.
- InferencePost-training quantizationQuantization applied after training to reduce inference memory and cost.
- ConceptPrecision and RecallMetrics for classification: Precision is correct positives / predicted positives; Recall is correct positives / actual positives.
- Evaluation & safetyPrecision-recall AUCA summary of the precision-recall relationship across thresholds, particularly informative with a rare positive class.
- ConceptPredictionThe output produced by a trained model when given new input data.
- FoundationsPredictive uncertaintyUncertainty associated with an individual prediction, useful for abstention, ranking and risk control.
- InferencePrefix cachingReusing computation results for an identical prefix shared across requests.
- RepresentationPrincipal component analysisA linear method projecting data onto orthogonal directions of maximum variance to reduce dimensions and redundancy.
- ProbabilityPrior and posteriorDistributions representing parameter knowledge before data and updated knowledge after observations, respectively.
- Evaluation & safetyPrivacy budgetQuantification of cumulative privacy consumption in differentially private mechanisms, often expressed through epsilon.
- ProbabilityProbabilityA numerical measure between zero and one that assigns plausibility to events under coherent rules of addition and composition.
- RAG & retrievalProduct quantizationA technique that compresses vectors by splitting each embedding into subvectors and quantizing them separately.
- Agenti e protocolliProgressive disclosureGradual loading of information only when needed, starting from indices or descriptions rather than complete content.
- TechniquePrompt EngineeringThe practice of designing effective text prompts to guide large language models toward desired outputs.
- LLMs & promptingPrompt injectionMalicious instructions embedded in inputs or retrieved content to divert an AI system’s behavior.
- TechniquePruningRemoving unnecessary weights/neurons to reduce model size and computational cost.
Q
- TechniqueQuantizationReducing precision of weights/activations to lower memory and computation.
- InferenceQuantization-aware trainingTraining that simulates quantization effects to make the model more robust to reduced precision.
- RAG & retrievalQuery expansionEnriching a query with terms or reformulations to increase retrieval coverage.
- RepresentationQuery rewritingRewriting a request to make it self-contained, explicit or better suited to the retrieval system.
- ConceptQuery, Key, Value (QKV)Three vectors used in attention mechanisms to compute weighted combinations of input elements.
- TaskQuestion AnsweringNLP task where model extracts or generates answers to questions based on context.
- Modelli e prodotti AIQwenAn Alibaba model family including language, multimodal, coding and reasoning variants at different scales.
R
- TechniqueRAG (Retrieval-Augmented Generation)A technique that enhances LLM outputs by retrieving relevant information from external knowledge bases.
- LearningRandom forestAn ensemble of trees trained on random samples and feature subsets, aggregating predictions by vote or average.
- LearningRandom seedA value initialising a pseudo-random generator to make a sequence of apparently random operations reproducible.
- LLMs & promptingReActA strategy alternating reasoning and actions, allowing a model to observe results before proceeding.
- AI contemporaneaRealtime multimodal APIA low-latency bidirectional interface for streaming audio, text, video and tool calls.
- LLMs & promptingReasoning modelA model optimized to produce and evaluate complex intermediate reasoning steps before answering.
- Evaluation & safetyRecall@KThe share of relevant items retrieved within the top k results relative to all available relevant items.
- ConceptReceptive FieldThe region of input space that affects a particular neuron's activation in a neural network.
- Evaluation & safetyRed teamingDeliberately adversarial evaluation that seeks dangerous behavior, abuse and weaknesses in a system.
- TaskRegressionA supervised learning task where the goal is to predict continuous numerical values.
- TechniqueRegularizationTechniques to prevent overfitting by adding constraints or penalties to the model during training.
- Evaluation & safetyReidentificationAssociating a supposedly anonymous record with a person through combinations of attributes or external sources.
- ParadigmReinforcement LearningA machine learning paradigm where agents learn by interacting with an environment and receiving rewards or penalties.
- ConceptReLU (Rectified Linear Unit)An activation function that outputs the input if positive, otherwise zero: f(x) = max(0, x).
- RAG & retrievalRerankingReordering a limited candidate set with a more accurate model than the initial retrieval stage.
- ArchitectureResNet (Residual Network)A CNN architecture that uses residual connections (skip connections) to enable training of very deep networks.
- LLMs & promptingReward modelA model trained to score responses according to human preferences or learned criteria.
- LLMs & promptingRLHF (Reinforcement Learning from Human Feedback)Alignment that uses human preferences to guide a model through reinforcement learning.
- AlgorithmRMSpropAdaptive learning rate optimization algorithm using moving average of squared gradients.
- ArchitectureRNN (Recurrent Neural Network)A neural network architecture designed for sequential data, with connections that loop back to previous states.
- Evaluation & safetyRobustnessThe ability to maintain acceptable performance under noise, perturbations, shifts or adverse operating conditions.
- Evaluation & safetyROC curveA curve showing true-positive rate and false-positive rate as the classification threshold varies.
- Hardware e infrastrutturaROCmAn AMD GPU computing and machine-learning software platform, providing an alternative to CUDA on compatible hardware.
- Video e media generativiRunway Gen-4A generation of Runway video models focused on coherent characters, environments and controlled audiovisual production.
S
- ConceptScalarA single numerical value, a zero-dimensional tensor.
- ProbabilitySelective predictionA strategy allowing a model to abstain on uncertain cases and cover only a fraction of inputs.
- TechniqueSelf-AttentionAn attention mechanism used in deep learning models that allows a neural network to weigh the importance of different parts of an input relative to each other.
- ParadigmSelf-Supervised LearningLearning paradigm where models create supervision signal from unlabeled data.
- InferenceSemantic cachingA cache that retrieves answers or results based on semantic request similarity rather than exact equality alone.
- RAG & retrievalSemantic chunkingChunking that creates segments based on content meaning and structure rather than fixed length alone.
- TaskSemantic SegmentationA computer vision task that assigns a class label to every pixel in an image.
- RepresentationSentencePieceA subword tokenisation algorithm and implementation operating directly on text without requiring prior word segmentation.
- TaskSentiment AnalysisNLP task determining emotional tone or opinion expressed in text.
- LearningSequence-to-sequenceA paradigm transforming an input sequence into an output sequence through encoder-decoder or equivalent architectures.
- AlgorithmSGD (Stochastic Gradient Descent)A gradient descent variant that updates weights using gradients from a single random training example at a time.
- Evaluation & safetySHAPAn attribution method estimating each feature’s contribution using Shapley-value principles.
- ConceptSigmoid FunctionAn activation function that maps inputs to values between 0 and 1: f(x) = 1/(1 + e^(-x)).
- Generative AISim-to-realTransferring behaviors learned in simulation to systems operating in the real world.
- RepresentationSingular value decompositionA matrix factorisation into orthogonal components and singular values describing principal scales and directions.
- Agenti e protocolliSkill packageA package of instructions and resources providing an agent with a specialised procedure to load when relevant.
- RepresentationSliding window attentionAn attention scheme where each token attends only to a local window of nearby positions to reduce long-sequence cost.
- ConceptSoftmax FunctionAn activation function that converts a vector of values into a probability distribution summing to 1.
- Video e media generativiSoraAn OpenAI system for generating and transforming video from text instructions, images or other multimedia inputs.
- OptimisationSparse attentionA family of mechanisms computing only a subset of attention pairs to reduce memory and operations.
- RAG & retrievalSparse retrievalRetrieval based on sparse representations and term occurrences, such as inverted indexes and BM25.
- Audio e multimodaleSpeaker diarisationSegmentation of a recording assigning audio portions to different speakers without necessarily identifying them by name.
- RepresentationSpecial tokensReserved symbols marking start, end, separation, padding or structural roles within a tokenised sequence.
- InferenceSpeculative decodingDecoding that uses a smaller model to propose tokens and a main model to verify them in parallel.
- Audio e multimodaleSpeech-to-speechInteraction in which speech and audio are processed and returned without requiring a visible intermediate text transcription.
- DatasetSQuAD (Stanford Question Answering Dataset)Reading comprehension dataset with 100k+ questions on Wikipedia articles.
- Video e media generativiStable Video DiffusionA Stability AI model extending latent diffusion to short video-sequence generation from images.
- ProbabilityStandard deviationThe square root of variance, expressing dispersion in the same unit as the observed variable.
- RepresentationStandardisationA transformation subtracting the mean and dividing by standard deviation to produce centred, unit-scale features on training data.
- ConceptStrideThe number of pixels by which a filter moves across the input during convolution or pooling operations.
- LLMs & promptingStructured outputOutput generated according to a defined schema, such as JSON with required fields and types.
- ParadigmSupervised LearningA machine learning paradigm where models learn from labeled training data with input-output pairs.
- LearningSupport vector machineA model learning a boundary by maximising margin, optionally using kernels for nonlinear relationships.
- FoundationsSynthetic dataArtificially generated data from models, simulations or rules, used to augment or replace real observations.
- AI contemporaneaSystem 0A theoretical perspective on AI as an external cognitive environment shaping attention and decisions before deliberate human reasoning.
- AI contemporaneaSystem 1 and System 2 in AIA metaphor distinguishing fast associative responses from slower, deliberate and verifiable processes in AI systems.
- LLMs & promptingSystem promptHigh-priority instructions that define a model’s behavior, constraints and operating context.
T
- Video e media generativiTalking-head generationSynthesis of a speaking face by synchronising lips, expressions and motion with an audio track or text.
- RepresentationTarget encodingEncoding categories through target statistics estimated from data, often with smoothing for small groups.
- TechniqueTemperature (Sampling)Parameter controlling randomness in text generation by scaling logits before softmax.
- Video e media generativiTemporal consistencyThe property by which subjects, identities, lighting and geometry remain coherent across consecutive frames of generated video.
- Hardware e infrastrutturaTensor CoreAn NVIDIA hardware unit specialised in reduced-precision matrix operations used for accelerated training and inference.
- OptimisationTensor parallelismPartitioning operations and tensors within one layer across devices to distribute memory and computation.
- Hardware e infrastrutturaTensor processing unitA Google accelerator designed for tensor workloads in neural-network training and inference.
- InferenceTensorRTAn NVIDIA SDK that optimizes and runs deep learning models on compatible GPUs.
- Hardware e infrastrutturaTensorRT-LLMAn NVIDIA library for optimising LLM inference with kernels, precision and execution strategies for supported GPUs.
- ConceptTest DataData held out for final evaluation of a trained model, never seen during training.
- LLMs & promptingTest-time computeAdditional computation used during inference to improve a response without changing model parameters.
- TaskText SummarizationNLP task condensing long text while preserving key information.
- Generative AIText-to-imageGenerating images guided by a textual description.
- Video e media generativiText-to-videoGeneration of a video sequence guided by a textual description of subjects, actions, environment and style.
- TechniqueTokenizationThe process of breaking text into smaller units (tokens) like words, subwords, or characters for processing.
- Agenti e protocolliTool discoveryThe process by which an agent finds available tools and understands their descriptions, parameters and conditions of use.
- AgentsTool sandboxingIsolation of tools and their actions within restricted permissions, resources and environments.
- LLMs & promptingTool useA model’s ability to request external tools to consult data or perform actions during a task.
- TechniqueTop-K SamplingText generation technique sampling from only the K most likely next tokens.
- ConceptTraining DataThe subset of data used to train a machine learning model.
- TrainingTransfer LearningA technique where knowledge learned from one task is applied to a different but related task, reducing training time and data requirements.
- ArchitectureTransformerA neural network architecture based entirely on attention mechanisms, without recurrent or convolutional layers.
- LLMs & promptingTree of ThoughtsA problem-solving method that explores and evaluates multiple intermediate reasoning paths before choosing.
- TechniqueTriplet LossTriplet loss is a metric-learning objective for training embeddings. It compares an anchor with a positive sample from the same class and a negative sample from a different class.
- Generative AITTS (Text-to-Speech)Speech synthesis from text, with control over voice, prosody and pronunciation.
U
- ArchitectureU-NetA CNN architecture designed for biomedical image segmentation, featuring an encoder-decoder structure with skip connections.
- Hardware e infrastrutturaUnified memoryAn architecture in which processors and accelerators share memory space, avoiding separate copies between system and device memory.
- ParadigmUnsupervised LearningUnsupervised learning finds structure in data without target labels. Its objectives include grouping similar observations, compressing representations and estimating the data distribution.
V
- ArchitectureVAE (Variational Autoencoder)A generative model that learns a probabilistic latent space representation of data.
- ConceptValidation SetA portion of data held out during training to tune hyperparameters and prevent overfitting.
- ConceptVanishing GradientA problem in deep networks where gradients become extremely small, preventing effective learning in early layers.
- ProbabilityVarianceThe mean squared deviation from the expected value, used to measure the dispersion of a random variable.
- ConceptVectorAn ordered array of numbers representing a point in multi-dimensional space.
- RAG & retrievalVector databaseA database designed to store embeddings and retrieve similar items through vector indexes.
- RepresentationVector normA function assigning length or magnitude to a vector under positivity, homogeneity and triangle-inequality properties.
- Video e media generativiVeoA Google DeepMind video-model family for cinematic generation and controls over scenes, subjects and motion.
- ArchitectureVGGDeep CNN architecture using small 3x3 filters throughout, emphasizing depth.
- Generative AIVideo generationGenerating coherent video sequences from text, images or other conditions.
- Video e media generativiVideo inpaintingReconstruction of missing regions or object removal in video by propagating coherent content through space and time.
- Video e media generativiVideo understandingAutomatic analysis of content, actions, relationships and events in video sequences beyond single-frame recognition.
- Video e media generativiVideo upscalingIncreasing a video’s spatial or temporal resolution by reconstructing missing detail and frames with learned models.
- Video e media generativiVideo-to-videoTransformation of an existing video guided by prompts, references or conditions while preserving some original-sequence properties.
- Generative AIVision-language model (VLM)A model that jointly processes images and language to understand or generate multimodal content.
- Generative AIVision-language-action (VLA)A model connecting visual perception and language instructions to physical or robotic actions.
- ArchitectureViT (Vision Transformer)Transformer architecture adapted for computer vision by treating image patches as tokens.
- Hardware e infrastrutturavLLMAn LLM serving engine optimising throughput and memory with techniques such as PagedAttention and continuous batching.
- Generative AIVoice cloningSpeech synthesis that reproduces characteristics of a specific voice from audio samples.
W
- Video e media generativiWanAn Alibaba video-model series with variants for text, image and other forms of controlled generation.
- Evaluation & safetyWatermarkingEmbedding a detectable signal in content or models to attest origin or provenance.
- InferenceWebGPUA web API for GPU compute and graphics that can also run model inference in the browser.
- ConceptWeightLearnable parameters in neural networks that determine the strength of connections between neurons.
- FoundationsWeight decayA regularization technique that penalizes large parameter values during model optimization.
- TechniqueWeight InitializationMethods for setting initial values of neural network weights before training begins.
- Audio e multimodaleWhisperAn OpenAI model for speech recognition, multilingual transcription and translation into English from audio.
- TechniqueWord EmbeddingDense vector representations of words that capture semantic and syntactic relationships.
- RepresentationWordPieceA subword tokenisation method building a vocabulary to represent frequent whole words and others through fragments.
- Generative AIWorld modelA learned model of environment dynamics used to predict future states and action consequences.
Y
Z
No terms found.