Artificial Intelligence Glossary
227 terms
A
A* (A-star)
A* (A-star) is an informed search algorithm that finds the least-cost path in a graph by combining the cost already traveled with a heuristic estimate. We explain its formula f = g + h, when it guarantees the optimal path, its uses—from GPS to video games—and its main limit: memory.
Accuracy
Accuracy is the fraction of a classifier's correct predictions, but it misleads on imbalanced classes —hence it is paired with precision, recall, F1 and MCC.
Activation Function
The activation function is the piece that introduces nonlinearity into a neural network, and without it deep networks could not learn complex patterns. We explain why it is essential, the main functions—sigmoid, tanh, ReLU and softmax—with their virtues and problems, and why ReLU marked a turning point.
Active Learning
Active learning lets the algorithm choose which examples to learn from, asking a human annotator for the most informative labels to reach good accuracy with less labeled data. We explain its human-in-the-loop cycle, the strategies for deciding what to ask, its relation to semi-supervised learning, and the risk of sampling bias.
AdaBoost
AdaBoost combines weak classifiers, reweighting the misclassified examples each round to form a more accurate weighted vote. Its flip side is a sensitivity to noise and outliers that the multiclass (SAMME) and regression (AdaBoost.R2) variants do not remove.
Adaptive Growth Neural Networks
Architectures that do not fix their size in advance but add neurons or layers during training. The label is non-standard: the real family is constructive neural networks, from Cascade-Correlation to architecture search.
Affective Computing
Affective computing is the field of AI that seeks to recognize, interpret and simulate human emotions from the face, voice, physiology and text. Founded by Rosalind Picard (1997), it promises more responsive interfaces, though reliably reading emotion from the face remains scientifically contested.
Algorithms
What an algorithm is: a finite, well-defined sequence of steps for solving a problem or performing a calculation. Its properties according to Knuth, classic examples, complexity, and its relationship with artificial intelligence.
Alpha-Beta Algorithm
Alpha-beta pruning is an optimization of the Minimax algorithm that reduces the positions to evaluate in a game tree without changing the chosen move. We explain how it works with the alpha and beta values, how much compute it saves depending on move order, and where it is used, from chess to Go's handover.
Anomaly Detection
Anomaly detection spots the observations that deviate from what is «normal». We review its types, the label-based approaches and methods such as Isolation Forest and autoencoders, along with their challenges: class imbalance, drift and false positives.
Ant Colony Algorithm
The ant colony algorithm is an optimization metaheuristic inspired by the behavior of real ants foraging for food. We explain the ingenious biological idea behind it—the pheromone trail—how it carries that into problem-solving, its origin and its applications in route and network optimization.
Artificial Intelligence
Artificial intelligence is the discipline that seeks to build systems able to perform tasks that would normally require human intelligence. We explain a rigorous definition, its origin at the 1956 Dartmouth workshop, the distinction between narrow, general and superintelligent AI—and where we stand in 2026—and its major branches, from the symbolic approach to deep learning and generative AI.
Artificial Neural Networks
What an artificial neural network is: a model of weight-connected neurons that learns from data, inspired by the brain but not a simulation of it. It covers the neuron, the layers, training by backpropagation and the limits.
Attention
The mechanism that lets a neural network dynamically weigh the relevant parts of its input instead of compressing it into a fixed vector. From Bahdanau's 2014 machine translation to the Transformer, it is now the foundation of large language models.
Attention Mechanisms
What attention mechanisms are, how Bahdanau, Cho and Bengio (2014) introduced them in machine translation, and why the Transformer's query-key-value scheme and multi-head attention made them the core of large language models.
AUC-ROC
AUC-ROC is the area under the ROC curve, summarizing in a single number a binary classifier's performance across all thresholds. We explain its probabilistic reading and three often-forgotten limits: it measures discrimination but not calibration, sets no threshold, and can be optimistic under class imbalance.
Autoencoders
A neural network that learns by squeezing its input through a bottleneck and rebuilding it, uncovering the essential structure of the data. From dimensionality reduction to anomaly detection and the generative VAE.
B
Backpropagation
Backpropagation is the algorithm that computes the gradient of the loss function in a neural network via the chain rule. We clear up a common confusion—computing the gradient is not updating the weights, which is the optimizer's job—explain the forward and backward passes, and review its history.
Bagging
Bagging is an ensemble learning technique that trains several models on different random samples of the dataset and combines their predictions. Its main effect is to reduce variance and overfitting. We explain its bootstrap-and-aggregate mechanics, why it works, its link to Random Forest and its contrast with boosting.
Batch Gradient Descent
Batch Gradient Descent computes the gradient over the entire training set before each update: its steps are exact and stable, but slow and hard to scale next to stochastic and mini-batch gradient descent.
Bayes' Theorem
Bayes' theorem is the rule for updating our beliefs in light of evidence, blending what we already knew with what we have just observed. It underpins Bayesian inference and the Naive Bayes classifier.
Bayesian Inference
Bayesian inference is a statistical method that updates the probability of a hypothesis as new evidence arrives, via Bayes' theorem. We explain its three elements—prior, likelihood and posterior—why it is not a synonym for artificial intelligence, and what it is used for in machine learning.
Bayesian Networks
A Bayesian network is a probabilistic graphical model that represents random variables and their dependencies with a directed acyclic graph. We explain how it compactly encodes the joint distribution with conditional probability tables, what inference means, and its uses, from medical diagnosis to spam filtering.
BERT
BERT, the bidirectional encoder Google unveiled in 2018, taught machines to read language from context. We cover how it is trained with MLM and NSP —and why RoBERTa showed NSP was dispensable—, its place beside generative LLMs, and the encoders that update it today.
Bias
«Bias» names at least four distinct concepts in AI: a neuron's intercept, an estimator's bias, the bias of the bias-variance tradeoff, and algorithmic bias. This entry pulls them apart one by one.
Big Data
Big Data is neither an era nor a single technology, but the problem of extracting value from data that overwhelms conventional tools. We cover Laney's V's, the shift from Hadoop and Spark to the cloud and the lakehouse, data governance, and its real —and sober— relationship with AI.
Boltzmann Machines
A Boltzmann machine is a recurrent, stochastic, generative, energy-based neural network. We explain its structure of visible and hidden units, its practical version—the restricted machine (RBM)—and why, despite its enormous historical influence, it is now largely in disuse compared with other generative architectures.
Boosting
Boosting chains many weak learners in sequence —each focused on the previous one's errors— into a strong predictor. From AdaBoost to the gradient boosting of XGBoost, LightGBM, and CatBoost, it remains the go-to technique for tabular data.
C
Case-Based Reasoning
Case-Based Reasoning solves new problems by reusing solutions from similar past cases, through the four-R cycle (retrieve, reuse, revise, retain) formalised by Aamodt and Plaza in 1994.
CatBoost
CatBoost is a Yandex gradient boosting library that handles categorical features natively. We explain its contributions—ordered boosting and symmetric trees—what target leakage is and how it corrects it, and when to prefer it over XGBoost and LightGBM.
Classification
Classification assigns each input to a discrete category through supervised learning, unlike regression, which predicts a number. We distinguish its types (binary, multiclass and multilabel), the real range of algorithms, and how it is evaluated when classes are imbalanced.
Clustering
Clustering splits unlabeled data into groups by similarity. A tour of its main families —k-means, hierarchical, DBSCAN and Gaussian mixtures—, how to choose the number of groups and why there is no single correct answer.
Collaborative Filtering
Collaborative filtering is the recommendation technique that predicts what a user will like from the behavior of many others. We explain its principle, its contrast with content-based filtering, its neighborhood and matrix-factorization approaches—popularized by the Netflix Prize—its known problems and its neural version.
Collaborative Learning
In artificial intelligence, collaborative learning is the paradigm in which several entities—models, devices or nodes—train jointly by sharing knowledge rather than data. We clarify its two senses, place it as an umbrella with federated learning at its center, distinguish it from collaborative filtering, and review its benefits and challenges.
Community Detection
Community detection is the task of finding, within a network, groups of nodes that are highly connected to each other and less linked to the rest. We explain the concept of modularity, the classic algorithms (Girvan-Newman, Louvain, Leiden, label propagation), the challenges such as the resolution limit and overlapping communities, and its current relation to AI.
Computer Vision
How AI teaches machines to interpret images and video: from hand-crafted features to AlexNet's 2012 leap, Vision Transformers and multimodal models, with their still-open uses and limits.
Conditional Generative Models
A conditional generative model generates data conditioned on additional information, modeling p(x|y) instead of p(x), which allows controlling what is produced. We explain its contrast with unconditional models, how it is implemented in GANs, variational autoencoders and diffusion, and its uses.
Constraint Satisfaction
A constraint satisfaction problem (CSP) is defined by variables, their domains of possible values and the constraints that limit valid combinations. We explain what a solution is, examples such as Sudoku or scheduling, the methods to solve them, and the difference between satisfaction and constrained optimization.
Content-Based Filtering
Content-based filtering is the recommendation approach that suggests items similar to those the user already rated well, looking at the items' own attributes and not at other users' behavior. We explain how it builds the user profile, its contrast with collaborative filtering, and its advantages and limits.
Continuous Neural Networks
«Continuous neural networks» is not a method with a name of its own but the continuous-depth, continuous-time family whose most verifiable form is Neural ODEs: they replace the stack of discrete layers with a transformation governed by a differential equation.
Convolutional Neural Networks
The deep learning architecture that learns to see images through filters sliding across them, with shared weights and translation invariance. From LeNet-5 to AlexNet, CNNs unlocked modern computer vision.
Cosine Distance
Cosine similarity measures the cosine of the angle between two vectors, and cosine distance is 1 minus that similarity. We explain the difference between them, that they measure orientation and not magnitude, their use for comparing text embeddings, and a technical caveat: cosine distance is not a metric in the strict sense.
Cosine Restart (Warm Restarts)
Cosine restart, or cosine annealing with warm restarts (SGDR), is a strategy for scheduling the learning rate during training: it combines a cosine-shaped decay with periodic restarts that help escape local minima. We explain how it works, why it is used, and its parameters.
Cross-Entropy
Cross-entropy measures, in bits, how far a prediction departs from reality; it decomposes into entropy plus Kullback-Leibler divergence and is the standard loss function for classification in machine learning.
Cross-Validation
Cross-validation is a resampling technique that estimates how a model will generalize to unseen data by reusing what is available. Its flagship method, k-fold, is used to compare models, tune hyperparameters and detect overfitting.
Curriculum Learning
Curriculum learning orders training examples from easy to hard. We look at how that difficulty is measured and why the evidence is mixed: it helps with noisy data or tight budgets, but not always.
D
Data Association
Data association is the problem of deciding which sensor observation belongs to which object when several are being tracked at once, amid noise and false alarms. We explain why it is hard, the classic methods (nearest neighbor, JPDA, multiple hypothesis tracking), its relation to the Kalman filter and its modern version with deep learning.
Data Compression
The technique that reduces the bits needed to represent information. Its great divide separates lossless compression —bounded by Shannon entropy— from lossy compression, and today it links to AI: compressing well means predicting well.
Data Integration
Data integration combines heterogeneous sources —databases, APIs, files and sensors— into a single, coherent view, reconciling schemas, entities and quality. It is the precondition for any AI model to learn from trustworthy data.
Data Mining
Data mining is the process of discovering useful, non-obvious patterns in large volumes of data, at the intersection of statistics, machine learning and databases. We explain its relation to the KDD process, its main tasks, the CRISP-DM methodology, its difference from neighboring terms and its privacy considerations.
Decision Trees
A model that predicts by chaining simple questions about the data through recursive partitioning. Its strength today lies in ensembles — Random Forest, XGBoost, LightGBM, CatBoost — which still dominate tabular data against deep learning.
Deep Learning
Deep learning uses many-layered networks to learn hierarchical representations of data, without hand-designed features. We review its fundamentals (backpropagation, representation learning), its architectures and scaling, and its limits: data, compute, robustness, interpretability, hallucinations and bias.
Deep Neural Networks
A deep neural network has several hidden layers; “deep” refers to the number of layers. We clarify that it is not a subdiscipline but a class of models, and that more depth does not imply better performance: it depends on architecture, data and task. We review its architectures and the trend toward efficiency.
Deep Reinforcement Learning
Deep Reinforcement Learning pairs reinforcement learning with deep neural networks that approximate the policy or the value, enabling decisions from high-dimensional states such as pixels. From DQN to AlphaGo, its milestones, its methods (value, policy, PPO) and its honest limits.
Dialogue Systems
What a dialogue system, or conversational agent, is: from the task-oriented versus open-domain divide and the classic NLU–manager–NLG pipeline to ELIZA and the large-language-model turn, honest limits included.
Dimensionality Reduction
What dimensionality reduction is and why selection, extraction, and visualization are different jobs: from PCA to UMAP and embeddings, plus the traps of data leakage and misread maps.
Dirichlet Processes
A Dirichlet process is a distribution over distributions and a cornerstone of Bayesian nonparametric statistics. It lets you cluster data without deciding the number of groups in advance: the model infers it and lets it grow with the data.
Discriminative Models
A discriminative model learns to separate classes by modeling p(y|x), unlike the generative model, which models the joint distribution p(x,y) and can generate data. We clarify that it is not the same as a “classifier,” with examples of each approach, and what suits which data regime (Ng and Jordan, 2001).
DistilBERT
DistilBERT is a compact version of BERT built by knowledge distillation: per its 2019 paper, about 40% smaller, 60% faster and retaining 97% of performance. We explain how distillation works, why it is not «pruning» the model, and where it stands against today's compression toolkit.
Distributed Artificial Intelligence
The AI subfield where several distributed agents or nodes cooperate or compete to solve problems beyond any single one: from distributed problem solving and multi-agent systems to the 2026 revival with LLMs and federated learning.
Dynamic Belief Systems
«Dynamic Belief Systems» is not a standard technical term: it describes how an agent maintains and updates a probabilistic belief —a distribution over the world— as new evidence arrives. Its verifiable basis is dynamic Bayesian networks and Bayes filtering.
E
Ensemble Learning
Ensemble learning combines the predictions of several models to obtain a better, more robust result than any of them alone: the “wisdom of the crowd” applied to machine learning. We explain why it works, its three main methods—bagging, boosting and stacking—and its advantages and limits.
Entity Extraction
Entity extraction (NER) locates and classifies mentions such as people, places or dates in a text. We distinguish it from relation extraction and entity linking, trace the methods from rules to language models—with the right nuance about BERT—and its privacy dimension.
Entropy
Shannon entropy measures the uncertainty or average information of a random variable. We give its definition and formula, explain its intuition in bits, distinguish entropy from cross-entropy, and survey its uses in machine learning, from decision trees to reinforcement learning.
Ethics in Artificial Intelligence
Ethics in artificial intelligence studies the moral implications of designing, developing and deploying AI systems, and proposes principles so they are beneficial and respect people's rights. We explain those recurring principles, the major international frameworks (UNESCO, EU, OECD), the concrete problems they address and why it is an open debate.
Euclidean Distance
Euclidean distance is the straight-line separation between two points, a generalization of the Pythagorean theorem. We explain its use as a similarity measure in machine learning (k-NN, k-means), two important cautions—scale and dimensionality—and its relation to other metrics.
Evaluation Metrics
Evaluation metrics are the quantitative measures used to judge a model, and there is no single one. We cover the classification metrics (confusion matrix, precision, recall, F1 and ROC-AUC), why accuracy misleads on imbalanced classes, and the regression metrics (MAE, MSE, RMSE and R²).
Event Detection
Event detection identifies occurrences of interest in data streams —time series, text, audio or video—; unlike an anomaly, an event is a recognizable happening that may be frequent and expected.
Evolutionary Learning
Evolutionary learning is an optimization and machine learning approach inspired by biological evolution: it evolves a population of solutions across generations. We explain its mechanics, the evolutionary computation family, how it applies to learning—including neuroevolution—its advantages over gradient methods and its limits.
Expert Systems
What an expert system is: an AI program that emulates a human specialist's judgement through a knowledge base of IF-THEN rules and an inference engine. Its golden age with DENDRAL and MYCIN, and why machine learning displaced it.
Explainability and Transparency in AI
Explainability, interpretability and transparency are not the same: explaining an individual decision, building a comprehensible model and documenting its governance are three distinct demands, each with its own methods, regulation and limits —including explanations that may not be faithful to the model.
F
F1 Score
The F1 score combines precision and recall through their harmonic mean: F1 = 2 · (precision · recall) / (precision + recall). We cover its micro, macro and weighted variants, how it compares with accuracy and MCC, and its limits.
FastText
FastText, Facebook AI Research's open-source library that represents every word as a bag of character n-grams. That lets it build vectors for words it never saw and excel at morphologically rich languages and fast text classification.
Feature Engineering
Feature engineering turns raw data into variables a model can use. Central to classical machine learning, deep learning has reduced —not erased— it by learning representations on its own.
Feature Selection
Feature selection keeps a subset of the original variables —discarding the rest while preserving interpretability— and differs from extraction, which builds new ones. Its biggest risk is data leakage: selecting before splitting train and test inflates performance.
Federated Learning
Federated learning trains a shared model across many devices without gathering their data on a server: each client trains locally and sends only model updates. We explain its client-server cycle, its uses and its limits—including that it does not by itself guarantee privacy.
Few-Shot Learning
Few-shot learning lets a model tackle a new task with very few labeled examples. We distinguish its two senses—classic meta-learning, where the examples train the model, and the in-context learning of large models, where they go in the prompt—and explain its methods and limits.
Finite-State Machine Learning
Learning a finite-state machine is not about using a hand-designed automaton but about inferring one from data. We cover what learning means, why it is hard (Gold, Angluin) and its real, and limited, relation to neural networks.
FrameNet
FrameNet is a hand-built lexical resource for English, created at ICSI Berkeley from Charles Fillmore's frame semantics, that describes words through the situations they evoke and the participants involved. A cornerstone of semantic role labeling, it now coexists with large language models.
Fraud Detection
How machine learning tells the few fraudulent operations from millions of legitimate ones: supervised classifiers versus anomaly detection, the asymmetric cost of false positives and human review, and why generative AI and concept drift force models to be validated without pause.
Fuzzy Logic
Fuzzy logic is a many-valued logic in which truth is a degree between 0 and 1, rather than the sharp true/false of classical logic. It serves to model vague concepts such as “tall” or “hot.” We explain its origin, fuzzy sets and membership functions, inference systems and its applications in control.
G
Gaussian Process Classification
Gaussian process classification applies a Bayesian non-parametric model—a distribution over functions—to the problem of assigning classes. We explain what a Gaussian process is, how it is adapted from regression to classification via a link function, why inference requires approximations, and what its advantages and main limit are.
Gaussian Process Regression
Gaussian process regression is a non-parametric Bayesian method that defines a probability distribution over functions and updates it with data. We explain the role of the kernel, its key advantage—predicting with a measure of uncertainty—and its main limit, a cost that scales with the cube of the number of data points.
Gaussian Processes
A Gaussian process is a distribution over functions, defined by a mean and a kernel, that underpins two tasks: regression with calibrated uncertainty and approximate classification. Its main limitation is the O(n³) cost, which motivates sparse approximations.
Generative Adversarial Networks
What Generative Adversarial Networks (GANs) are: two networks competing in a minimax game, their real limits —instability and mode collapse— and why diffusion models have displaced them by 2026.
Generative Models
What generative models are: systems that learn the distribution of data to create new samples —text, images, audio— rather than just classify them. We cover their four families (autoregressive, VAE, GAN and diffusion), why they power generative AI, and where their limits lie.
Genetic Algorithms
An evolution-inspired optimization metaheuristic: a population of solutions refined through selection, crossover and mutation across generations. Useful in hard search spaces, but with no guarantee of a global optimum and a high evaluation cost.
Gibbs Sampling
Gibbs sampling is a Markov chain Monte Carlo (MCMC) algorithm that draws samples from a difficult joint distribution by breaking it into simpler samplings: it updates each variable from its conditional distribution. We explain how it works, its origin, its uses in Bayesian inference and its convergence limits.
GloVe
GloVe turns words into vectors by starting from the global co-occurrence matrix of an entire corpus. A Stanford method that popularized the «king − man + woman ≈ queen» arithmetic and now stands as a landmark of static embeddings.
GPT
GPT (Generative Pre-trained Transformer) is not a model but a family of OpenAI systems built on a decoder-only transformer and pre-trained in a self-supervised way. A tour of its method and lineage, from GPT-1 to today's family, without crowning any single version.
Gradient Descent
Gradient descent is the algorithm that trains most deep learning models: it minimizes a function by moving, step by step, in the direction opposite to its gradient. We explain its mechanics, the crucial role of the learning rate, its variants by how much data they use, extensions such as Adam and its limits.
GRU
The GRU (gated recurrent unit) is a type of unit for recurrent neural networks that uses “gates” to control what information to keep over time. We explain its origin, how its two gates work, its contrast with the LSTM—simpler, with fewer parameters—and its place today, after the arrival of transformers.
H
Hamming Distance
Hamming distance counts how many positions two equal-length strings differ in. Born with Hamming's 1950 error-correcting codes, it is now key to comparing binary hashes and nearest-neighbor search in AI.
Handwriting Recognition
What handwriting recognition is, online and offline: from LeNet and CTC-trained LSTMs to transformers like TrOCR and Donut and on to multimodal models — how it is measured with CER and WER, and where its real limits lie.
Heuristic Search
Heuristic search uses a function that estimates the remaining cost to the goal to guide the exploration of a state space. We explain, with the correct notation, greedy search and the A* algorithm (f = g + h) and admissibility, and distinguish informed search from optimization metaheuristics.
Hidden Markov Models
A hidden Markov model (HMM) describes sequences in which hidden states generate visible observations. We explain what defines it (states, transitions, emissions, initial distribution), the three classic problems it solves (evaluation, Viterbi and Baum-Welch), and its place versus neural networks.
Hierarchical Clustering Algorithms
Hierarchical clustering is a family of unsupervised grouping methods that organizes data into a tree of nested clusters, the dendrogram. We explain its two approaches—agglomerative and divisive—how the dendrogram is read to choose the number of groups, the linkage criteria (single, complete, average and Ward) and its advantages and limits.
Hill Climbing Algorithm
Hill climbing is a greedy local search that always steps to the best neighbor until it reaches a peak. Fast and memory-light, its weakness is local optima, which variants such as random restart and simulated annealing help escape.
Huffman Coding
Huffman coding is a classic lossless compression algorithm from information theory —not an AI technique— that assigns shorter prefix codes to frequent symbols by building an optimal binary tree.
Human-Robot Interaction
Human-robot interaction (HRI) is the interdisciplinary field that studies how people and robots communicate, collaborate and work together. We explain what distinguishes it from interacting with a computer, the axes that characterize it, key concepts such as safety, trust and the uncanny valley, and its applications.
Hyperparameters
Hyperparameters are the settings fixed before training a model that control how it learns, unlike the parameters the model learns from data. We explain that distinction, the methods to tune them (grid, random, Bayesian), and why their choice requires validation to avoid overfitting.
I
ICA (Independent Component Analysis)
Independent component analysis (ICA) separates a signal made of several mixed sources into its original components, assuming they are statistically independent. We explain the cocktail party problem that illustrates the idea, how it differs from PCA, and its uses, from audio separation to biomedical signals.
Image Recognition
Image recognition is the computer-vision task that identifies and classifies the content of an image. We explain its nuances—classify, detect, segment—the turning point that AlexNet marked in 2012, how convolutional networks work in broad strokes, their applications and their limits, such as biases and adversarial examples.
Image Segmentation
To segment an image is to label every one of its pixels. This entry distinguishes semantic, instance, and panoptic segmentation, and traces the path from U-Net and Mask R-CNN to transformers and the SAM and SAM 2 foundation models, with their metrics, uses, and limits.
Importance Sampling
Importance sampling estimates hard averages by drawing from a convenient distribution q and reweighting each sample with the ratio p(x)/q(x). A good q slashes variance; a bad one makes it explode.
Inception
Inception is a family of convolutional neural network architectures for computer vision whose first version, GoogLeNet, won the 2014 ImageNet challenge. We explain its central innovation—the Inception module—the trick that made it efficient, its origin and its later evolution.
Incremental Learning
Incremental learning learns from sequential data without forgetting; its barrier, catastrophic forgetting, is only mitigated, never eliminated. Continual and lifelong learning name the same problem.
Inductive Logic Programming
Inductive logic programming is a branch of symbolic AI that learns logic rules from examples and background knowledge. Unlike statistical machine learning, it produces interpretable hypotheses and learns from very little data. We explain its idea, what sets it apart, its systems and its role in neuro-symbolic AI.
Information Extraction
Information extraction automatically obtains structured data—entities, relations, events—from unstructured text. We clarify that it is an umbrella of several subtasks and not a synonym for NER, trace the evolution of its methods from rules to large language models, and its challenges.
Information Theory
The mathematical framework Claude Shannon founded in 1948 to quantify information —entropy, mutual information, KL divergence and channel capacity— now at the heart of machine learning.
Instance-Based Learning
Instance-based learning, or “lazy learning,” is a family of methods that do not build a model during training but store the examples and respond by comparing each new query with them. We explain its canonical example—k-nearest neighbors—its contrast with “eager” learning and its advantages and limits.
Intelligent Agents
What an intelligent agent is: the entity that perceives its environment and acts to achieve goals, from Russell and Norvig's rational agent to today's large-language-model agents.
Intelligent Control
Intelligent control applies AI techniques —fuzzy logic, neural networks, adaptive control, and reinforcement learning— to govern dynamic systems where classical, model-bound control falls short. It does not replace it: it complements it where uncertainty rules.
J
K
L
LDA (Latent Dirichlet Allocation)
LDA (Latent Dirichlet Allocation) is the probabilistic topic model introduced by Blei, Ng, and Jordan in 2003: each document as a mixture of topics, each topic as a distribution over words. We explain how it works, its documented limits, and where it stands today against embeddings and large language models.
Learning Rate
The learning rate sets the step size of gradient descent: too high and training diverges, too low and it crawls. We cover its role, schedules like warmup and cosine annealing, and adaptive methods such as Adam.
Levenshtein Algorithm
The Levenshtein distance measures how different two text strings are: the minimum number of single-character edits—insert, delete or substitute—to turn one into the other. We explain its definition, its origin, how it is computed by dynamic programming, its properties and its uses, from spell-checkers to bioinformatics.
LightGBM
LightGBM is Microsoft's gradient boosting library, training decision trees with histograms and leaf-wise growth to deliver models that are accurate, fast and lightweight. We cover its GOSS and EFB techniques, its overfitting risk and how it compares with XGBoost and CatBoost.
Loss Function
A loss function measures how wrong a model is, and training adjusts its parameters to minimize it. We explain why it is not the same as a metric or the objective, review the most common losses by task type, and the conditions it must meet.
LSTM
LSTMs are recurrent neural networks that Hochreiter and Schmidhuber devised in 1997 to overcome the vanishing gradient and remember long-range dependencies. A memory cell and three gates explain why they dominated sequence modeling before the arrival of transformers.
M
Machine Learning
Machine learning is the branch of AI that develops algorithms able to learn from data and generalize to unseen data. We review its operational definition, the three main paradigms and self-supervision, its relation to deep learning and AI, and its limits (overfitting, bias, data quality).
Machine Translation
From RBMT to statistics to neural networks: how machine translation reached the Transformer and large language models, how it is measured with BLEU and its limits, and why it still fails on low-resource languages.
Mahalanobis Distance
The Mahalanobis distance is a measure —not an algorithm— that gauges the separation between a point and a distribution while accounting for the covariance between variables, making it invariant to scale and correlation.
Manhattan Distance
Manhattan distance measures the gap between two points as the sum of the absolute differences of their coordinates, like a taxi crossing a street grid with no diagonal shortcuts. As the p = 1 case of Minkowski distance, it is robust to outliers and useful in high dimensions.
Markov Chains
A Markov chain is a stochastic process in which the next state depends only on the current one, not on prior history. Through its transition matrix and stationary distribution it underpins hidden Markov models, MCMC sampling and Google's PageRank.
Maximum Entropy
The maximum entropy principle (Jaynes, 1957) picks the most uncertain distribution consistent with what is known. We separate three ideas often confused—the principle, cross-entropy and regularization—and clarify why the “maximum entropy model” is logistic regression, not a GAN or a Transformer.
Maximum Entropy Method
The maximum entropy method is a principle of inference that, among all probability distributions consistent with what we know, chooses the one with the greatest entropy: the most “honest,” the one that assumes nothing extra. We explain Jaynes's principle, its intuition, its version in machine learning—the maximum entropy classifier—and its applications.
Maximum Likelihood Estimation
The method that picks a model's parameters by making the observed data most probable: from Fisher to the log-likelihood and the cross-entropy that trains neural networks.
Meta-learning
Meta-learning—“learning to learn”—seeks to have a system improve its own learning process from experience across many tasks, so as to adapt quickly to new tasks with very few examples. We explain its relation to few-shot learning, its three major families of methods, its difference from transfer learning and its echo in large language models.
Metropolis-Hastings Sampling
An MCMC algorithm that samples hard distributions by proposing candidates and accepting them with a calculated probability. A foundation of Bayesian inference, it is distinct from HMC and NUTS, which use gradients.
Minimax Algorithm
In an adversarial game, the Minimax algorithm chooses the move that maximizes the guaranteed minimum gain against an optimal opponent. We explain how it traverses the game tree, its assumptions (two players, zero-sum, perfect information), alpha-beta pruning, and its—nuanced—relation to Nash equilibrium.
Minkowski Distance
The Minkowski distance is the general metric that, depending on the order p, becomes Manhattan (p = 1), Euclidean (p = 2) or Chebyshev (p → ∞).
Model Ensemble
Combining several models whose errors are diverse yields more reliable predictions than any single one. Bagging, boosting and stacking are the three families that exploit this idea.
Model Evaluation
Model evaluation measures how well a model generalizes to new data, not how it reproduces the training data. We explain the split into training, validation and test, cross-validation, task-specific metrics, and key concepts such as the bias-variance trade-off and data leakage.
Model Interpretation
Model interpretation gathers three ideas that are not synonyms: interpretability, explainability and causality. We explain why Shapley values (SHAP) attribute importance but do not prove causes, organize the methods (intrinsic and post hoc, local and global), and their limits of faithfulness and stability.
Model-Based Reasoning Systems
Model-based reasoning is a symbolic-AI method that diagnoses faults by comparing what is observed against what an explicit model of the system predicts. The clash between model and observation yields conflicts, and from them the diagnoses, including faults never seen before.
Models
A model is what you get by training an algorithm on data: the tuned parameters that make predictions. We untangle the classic model-versus-algorithm confusion and walk through model types, fit, and Box's maxim.
Modular Neural Networks
A modular neural network splits the problem into specialized subnetworks whose outputs are combined. We separate modularity from ensembles, cascades and the mixture of experts (MoE)—today key to scaling large language models—and distinguish the axes of architecture, routing and composition.
Multi-Agent Systems
A multi-agent system is made up of several autonomous agents that interact in a common environment to solve problems beyond a single agent. We first define what an agent is, review the system's properties (autonomy, decentralization, interaction), coordination among agents, and its uses, including systems based on large language models.
Multi-Domain Learning
Multi-domain learning trains a single model to solve the same task across several domains at once. We clearly distinguish it from multi-task learning, transfer learning and continual learning—where the EWC technique actually fits—and explain its methods.
Multi-Label Learning
Multi-label learning assigns several labels at once to each example, unlike multi-class classification, which picks just one. We walk through its approaches — from Binary Relevance to transformers with sigmoid outputs — its metrics, and the role of decision thresholds.
Multi-Objective Optimization
Multi-objective optimization means optimizing several conflicting objectives at once, where improving one usually worsens another. Instead of a single solution, it delivers a set of trade-offs. We explain Pareto dominance and the Pareto front, the methods to tackle it—from the weighted sum to NSGA-II—and why the final choice remains human.
Multi-Task Learning
Multi-Task Learning trains a single network to solve several related tasks at once by sharing representations among them. Matched well it improves generalization; matched poorly, it causes negative transfer.
Multi-View Learning
Multi-view learning combines several «views» —distinct feature sets— of the same data to learn better and exploit unlabeled examples. From Blum and Mitchell's 1998 co-training to the bridge with multimodal learning.
Multilayer Perceptron
The multilayer perceptron is a feedforward network with hidden layers and nonlinear activations, trained by backpropagation. We correct a widespread error—Hinton did not create it; its roots are in Rosenblatt (1958)—clarify that “multilayer” is not “deep,” and show it lives on inside Transformers.
Multiple-Instance Learning
Multiple-Instance Learning groups examples into «bags» and labels only the whole bag, not each instance: a form of weak supervision that makes it cheap to train models where annotating every data point would be unfeasible, from drug screening to spotting cancer in digital biopsies.
N
Natural Language Processing (NLP)
NLP is the branch of AI that enables machines to understand and generate human language. From rule-based systems to foundation models: what it covers, how it is evaluated, and its documented risks.
NMF (Non-negative Matrix Factorization)
Non-negative matrix factorization (NMF) decomposes a non-negative data matrix into the product of two smaller, also non-negative matrices. We explain why that constraint yields interpretable parts-based representations, its uses—from topic modeling to audio separation—and how it is computed.
O
One-Shot Learning
One-shot learning recognizes a new class from a single example, relying on metric learning (siamese, matching, prototypical networks) and meta-learning; in LLMs, «one-shot» also means giving a single example in the prompt (in-context learning).
Online Learning
Online learning updates the model with each item of a stream, rather than training once on a fixed set. We correct its theoretical basis—regret minimization, not the universal approximation theorem—distinguish online, incremental and continual learning, and explain concept drift.
Ontology Learning
Ontology learning builds ontologies semi-automatically from text, extracting concepts, hierarchies and relations. We explain what an ontology is, how this learning is organized in layers, and distinguish it from ontology engineering, relation extraction and knowledge graphs.
OpenAI
AI research organization founded in San Francisco in 2015. Its stated mission is to ensure that artificial general intelligence benefits all of humanity; it is the creator of GPT, ChatGPT, DALL·E, Whisper and Sora.
Opinion Mining
Opinion mining is, in essence, another name for sentiment analysis, but it usually emphasizes the structured extraction of opinions: not only whether a text is positive or negative, but about what, by whom and when. We explain Bing Liu's quintuple model, its tasks and its relation to aspect-based analysis.
Optical Character Recognition (OCR)
OCR turns text inside images into machine-editable text. We explain how it works—from the classic pipeline to Transformers like TrOCR and Donut—why BERT is not an OCR model, and how its accuracy is measured.
Optimization
Optimization is finding the best solution of an objective function, its maximum or minimum. In machine learning it is the engine of training: adjusting a model's parameters to minimize its error. We explain what an objective function is, the difference between convex and non-convex problems, gradient descent and its variants, and why the global optimum is not always reached.
Overfitting
Overfitting is one of the central problems of machine learning: it occurs when a model fits the training data so well—including its noise—that it generalizes poorly to new data. We explain how it is detected, its contrast with underfitting and the bias-variance trade-off, and the techniques to mitigate it.
P
Particle Swarm Optimization
An optimization metaheuristic by Kennedy and Eberhart (1995) inspired by flocks: each particle updates its velocity and position toward its personal best (pbest) and the swarm's best (gbest). It does not guarantee the global optimum and is sensitive to its parameters.
Partitional Clustering Algorithms
Partitional clustering splits data into K non-overlapping groups in a single pass, with K-means as its flagship method. We cover how it works, how to choose K, and why it struggles with irregularly shaped clusters.
Perceptron
The perceptron, introduced by Frank Rosenblatt in 1958, is the simplest model of an artificial neuron: a weighted sum followed by a step function. Its inability to solve XOR, pointed out by Minsky and Papert, opened the way to the multilayer perceptron.
Planning
Automated planning is the branch of AI that searches for a sequence of actions leading from an initial state to a goal. From the STRIPS formalism and the PDDL language to heuristic search and today's large-language-model agents.
pLSA
Introduced by Thomas Hofmann in 1999, pLSA gave topic modeling a probabilistic foundation, but its lack of a document-level generative model made it prone to overfitting; LDA fixed this in 2003 with a Dirichlet prior.
Poisson Processes
A Poisson process counts random events in time under precise assumptions: independence and a rate λ. We give its formal definition (Poisson counts, exponential times), distinguish the homogeneous from the non-homogeneous case, and explain why in neurons or markets it is a model with assumptions, not reality.
Pose Estimation
Pose estimation locates people's joints or an object's 6DoF position and orientation in images and video: how its 2D and 3D variants differ, how it is evaluated (PCK, OKS, MPJPE), and the models, uses, and risks defining the field today.
Precision
Precision is a classification metric that measures what proportion of the cases a model flags as positive really are. We explain its formula, its place in the confusion matrix, why it should not be confused with accuracy, its tension with recall and when a high precision is worth prioritizing.
Principal Component Analysis (PCA)
Principal component analysis (PCA) is the most common linear dimensionality-reduction technique: it transforms correlated variables into a new set of uncorrelated variables, ordered by the variance they explain. We explain how it works, why you should center and standardize the data, what it is used for and its limit: it only captures linear relationships.
Privacy and Security in AI
Privacy and security in AI are distinct problems with distinct defenses: protecting the data that feeds models, and protecting the model against attacks such as prompt injection. We map the real risks, why federated learning does not guarantee privacy on its own, and what the EU AI Act and the NIST framework already require.
Probabilistic Reasoning
Probabilistic reasoning is the AI approach to working with uncertain knowledge: it represents what is not known with probabilities and draws conclusions using the rules of probability. We explain its central tool, Bayesian networks, what inference means, and how it differs from logical reasoning.
PyTorch
PyTorch is the open-source deep learning framework born at Meta and governed since 2022 by the PyTorch Foundation under the Linux Foundation. From autograd to torch.compile and ExecuTorch: what sets it apart, how it has evolved since PyTorch 2.0, and when to choose it.
R
Random Forest
Random forest is one of the most widely used machine learning algorithms: it combines many decision trees, each trained with a touch of randomness, and aggregates their predictions. We explain its two sources of randomness, its origin, its notable advantages—robustness, feature importance—and its limits.
Recall
Recall—also called sensitivity or the true positive rate—is a classification metric that measures what proportion of the actual positive cases a model correctly identified. We explain its formula, its place in the confusion matrix, its tension with precision, the F1 score that combines them and when a high recall is worth prioritizing.
Recommender Systems
Recommender systems are information-filtering systems that predict which items will interest a user in order to suggest them: products, films, music or news. We explain their main approaches, the data they use, their classic problems, their evolution up to deep learning and the metrics by which they are evaluated.
Recurrent Neural Networks
Recurrent neural networks process sequences by maintaining a hidden state that acts as memory. We cover how they work, why the vanishing gradient led to the LSTM and GRU, and what role they still play alongside Transformers.
Regression
Regression is the supervised-learning task that predicts a continuous value —a number—, as opposed to classification, which predicts categories. From least-squares fitting to the MSE, MAE and R² metrics, regularized variants and the logistic-regression caveat.
Regularization
Regularization gathers the techniques that reduce a model's overfitting and improve its ability to generalize. We explain the classic L1 and L2 penalties and their Elastic-Net combination, other techniques specific to neural networks such as dropout, and the bias-variance trade-off they tune.
Reinforcement Learning
Reinforcement learning is the paradigm in which an agent learns by trial and error to maximise a cumulative reward within an environment. From Q-learning and MDPs to the RLHF that aligns today's large language models.
Representation Learning
Representation learning is the set of techniques that let a system discover for itself the useful features of raw data, instead of a human designing them by hand. We explain its core idea, its contrast with manual feature engineering, its forms—from embeddings to self-supervised learning—and why it is a pillar of deep learning.
ResNet
ResNet is the neural network architecture that introduced residual or “skip” connections, making it possible to train much deeper networks than before. We explain the degradation problem it solved, how its residual block works, its winning origin in 2015 and why its idea remains present even in Transformers.
Restricted Boltzmann Machines
Restricted Boltzmann machines are two-layer generative neural networks that, trained with contrastive divergence, helped spark the deep-learning revival around 2006. Today their value is mainly historical and pedagogical, having been displaced by models such as variational autoencoders, GANs and diffusion.
Robotics
Robotics is the interdisciplinary field devoted to the design, construction and use of robots: machines able to perceive their environment, decide and act physically on the world. We explain the sense-plan-act loop, its relation to artificial intelligence—which provides the intelligence—the origin of the words “robot” and “robotics,” and where it is heading in 2026.
Rule-Based Learning
The family of machine learning methods that induce readable if-then rules from data, unlike expert systems, whose rules are written by hand. From CN2 and RIPPER to interpretable scoring systems and neuro-symbolic AI.
Rule-Based Reasoning
Rule-based reasoning represents knowledge as if-then rules and uses an inference engine. We explain forward and backward chaining, why it is not the same as backpropagation or backtracking, and why these systems are not “self-evolving.”
S
Scikit-learn
Scikit-learn is an open-source machine learning library for Python, focused on classic machine learning. We review what it offers, its origin, the consistent API that made it a standard, and why it is not a deep learning framework.
Search Algorithms
In artificial intelligence, search algorithms solve problems by exploring a state space until reaching a goal. We explain what defines a search problem, the difference between uninformed and informed (heuristic) search, and other families such as local and adversarial search.
Self-Organizing Maps
Kohonen's Self-Organizing Maps project high-dimensional data onto a 2D grid while approximately preserving its topology: nearby points land on nearby neurons. Today t-SNE, UMAP and dedicated clustering are usually preferred, yet the SOM keeps uses of its own.
Semantic Segmentation
Semantic segmentation assigns a class label to every pixel of an image, producing a dense map of the scene. From fully convolutional networks (FCN) to DeepLab and Segment Anything, this guide sets it apart from classification, detection and instance segmentation.
Semi-Supervised Learning
Semi-supervised learning combines few labels with much unlabeled data. We explain when it works (smoothness, cluster and manifold assumptions), its methods (pseudo-labeling, consistency, graphs), why it is not the same as self-supervised learning, and its risk of confirmation bias.
Sentiment Analysis
Sentiment analysis, or opinion mining, is the natural language processing task that identifies and classifies the emotional charge of a text: whether it expresses a positive, negative or neutral opinion. We explain its levels of detail, the evolution of its approaches, its challenges—such as sarcasm or negation—and its applications.
Sequence Alignment
Sequence alignment arranges two or more DNA, RNA or protein sequences to identify regions of similarity. We explain its types (global and local), the classic dynamic-programming algorithms (Needleman-Wunsch, Smith-Waterman) and fast heuristics, and its relation to modern AI, from AlphaFold to edit distance.
Siamese Networks
A Siamese network uses two twin subnetworks with shared weights to measure how alike two inputs are in an embedding space. From signature and face verification to one-shot learning, it is the workhorse of metric learning.
Signal Processing
Signal processing is the classical engineering field that analyzes, filters, and transforms signals such as audio, images, and the ECG. AI leans on it to acquire and prepare data, even as deep learning increasingly learns from the raw signal itself.
Simulated Annealing
Simulated annealing is a probabilistic optimization metaheuristic inspired by the controlled cooling of metals. Its trick is to sometimes accept worse solutions to escape local optima, with a probability that falls as it cools. We explain its origin, acceptance criterion, uses and limits.
Social Network Analysis
Social network analysis studies social structures by representing them as graphs: people or organizations are nodes and their relationships are edges. We explain its centrality metrics, concepts such as communities and small-world, its historical roots and its current connection to AI through graph neural networks.
Spatial Data Analysis
Spatial data analysis studies data with a geographic location, where position conditions what is observed. We explain spatial autocorrelation (Tobler's first law of geography and Moran's I), concepts of its own such as the MAUP problem, and why it is not the same as a geographic information system.
Spatio-Temporal Data Analysis
Spatio-temporal data analysis handles observations tied to both a place and a moment, whose key trait is that they are almost never independent: spatial and temporal autocorrelation, irregular data, and temporal leakage in validation decide whether a model is trustworthy or merely looks it.
Speech Recognition
Speech recognition is the technology that converts spoken language into text. We explain what distinguishes it from related tasks, its evolution from hidden Markov models to end-to-end deep learning (wav2vec 2.0, Whisper), the metric it is evaluated with (the word error rate) and its outstanding challenges.
Speech Synthesis
Speech synthesis, or text-to-speech, is the technology that generates artificial speech from text. We explain its evolution from stitching recorded fragments to the neural models that now produce almost human voices, how the process is organized, the metric used to measure quality and the ethical dilemmas of voice cloning.
Spiking Neural Networks
Spiking neural networks (SNN) are biology-inspired models in which neurons communicate via discrete impulses in time. We explain how they work, the challenge of training them, and their efficiency on neuromorphic hardware, with the caveat that they do not replace conventional deep networks.
SSD (Single Shot Detector)
SSD (Single Shot MultiBox Detector) is an object detection architecture introduced by Wei Liu and colleagues at ECCV 2016 that locates and classifies objects in a single neural network pass, with no separate region-proposal stage.
Statistical Inference
What statistical inference is: drawing conclusions about a population from a sample through estimation, confidence intervals and hypothesis testing, with a correct reading of the p-value and of the frequentist and Bayesian frameworks.
Statistical Relational Learning
Statistical relational learning (SRL) combines first-order logic, which represents relationships between entities, with probabilistic models, which handle uncertainty. We explain the problem it solves—interconnected data that most machine learning ignores—its main formalisms, its relation to neighboring fields and its applications.
Stochastic Gradient Descent
Stochastic Gradient Descent (SGD) minimizes a model's error by estimating the gradient from a single example or mini-batch instead of the whole dataset. It is the engine of neural network training, from noise and the learning rate to variants such as momentum and Adam.
Supervised Learning
Supervised learning trains a model on labeled data—input and desired-output pairs—so it maps new inputs to their correct output. We explain its two tasks (classification and regression), how it works and why it seeks to generalize, and its contrast with other approaches, along with its limits: the cost of labeling and bias.
Support Vector Machines
Support vector machines look for the hyperplane that separates two classes with the widest margin, defined only by the points closest to the boundary. Soft margin, the C parameter and the kernel trick make them powerful on small, high-dimensional datasets.
T
t-SNE
t-SNE is a nonlinear dimensionality-reduction technique widely used to visualize high-dimensional data in two or three dimensions. We explain its purpose, how it works in broad strokes, the role of the perplexity hyperparameter and, very importantly, the caveats on how to read—and not read—its maps.
T5 (Text-to-Text Transfer Transformer)
T5, from Google (Raffel et al., 2019), treats every language task as a “text-to-text” problem. We explain that unified approach, its encoder-decoder architecture versus GPT and BERT, its pre-training on the C4 corpus, and variants such as mT5 and Flan-T5.
Temporal Data Analysis
Temporal data analysis exploits the fact that order in time is information. We define trend, seasonality and stationarity, explain why evaluation must respect temporal order, and survey models from ARIMA to foundation models such as TimesFM and Chronos.
TensorFlow
The open-source machine-learning library built by Google Brain and released in 2015. It works with tensors flowing through a computation graph and today shares the deep-learning throne with PyTorch.
Text Generation
Text generation, or natural language generation, automatically produces human text from an input. We explain how today's models do it—autoregressively, with sampling strategies—what it is for, how it is evaluated, and its risks, such as hallucinations and bias.
Text Mining
Text mining is the process of extracting useful information and patterns from large amounts of unstructured text, turning it into usable data. It combines natural language processing, statistics and machine learning. We explain its relation to data mining, the typical preprocessing, its tasks and the impact of large language models.
Time Series
A time series is a sequence of observations ordered in time —prices, temperature, demand— whose analysis aims above all to predict the future. We cover its components, stationarity and autocorrelation, methods from ARIMA to neural networks, and why validation must respect temporal order.
Topic Models
Topic models are unsupervised algorithms that discover the latent themes organizing a large collection of documents, without prior labels. We explain their core idea—a topic as a distribution over words—the classic model (LDA) and its antecedents, how they are interpreted and their evolution in 2026 with embeddings and transformers.
Transfer Learning
The technique of reusing knowledge a model learned on one task to solve a related one. It now underpins foundation-model adaptation through parameter-efficient fine-tuning (LoRA, adapters, prompt tuning), with gains that hinge on domain similarity — a poor transfer can hurt.
Transformers
Transformers are the neural network architecture behind today's large language models. Introduced in 2017, it replaced recurrence with the attention mechanism. We explain what self-attention is, how encoders and decoders are organized, its variants (BERT, GPT) and its main limit: the quadratic cost of attention.
U
U-Net
U-Net is the U-shaped convolutional architecture Ronneberger, Fischer and Brox created in 2015 to segment images from little data. Today it is also the backbone of the diffusion models that generate images.
Underfitting
Underfitting is when a model is too simple to capture the structure of the data, performing poorly on both the training and test sets—the hallmark of high bias. It is the opposite of overfitting and is cured with more capacity, better features, or less regularization.
Unsupervised Learning
Unsupervised learning finds structure in unlabelled data: clustering, reducing dimensions, estimating densities and learning representations. Its self-supervised form, which manufactures its own labels, is now the engine behind foundation models.
Unsupervised Representation Learning
Unsupervised representation learning lets a model learn useful representations of data by itself, without human labels. We clarify its relation to self-supervised learning, review its methods (autoencoders, contrastive, masked) and explain why it underpins foundation models.
V
Variance
Variance measures how far values spread around their mean —the expected value of the squared deviation— and, in machine learning, it also names a model's sensitivity to its training data.
VGG
VGG is a family of convolutional neural networks for images, created at Oxford in 2014, that demonstrated the value of depth by stacking small 3×3 filters. We explain its design idea, the VGG-16 and VGG-19 variants, its place in history—with the caveat that it is heavy and now surpassed—and its uses.
Video Analysis
Video analysis applies computer vision to image sequences over time. We separate its tasks (classification, detection, tracking, segmentation, temporal understanding), survey the models from 3D convolutions to video transformers and video-language systems, and its risks of bias and surveillance.
W
Weights
A weight is a number an AI model adjusts during training to reduce its errors. Millions or billions of those numbers, combined, are the only thing a model "learns" — nothing else is stored inside.
Word2Vec
Word2Vec (Google, 2013) learns word vectors from their contexts, via the CBOW and Skip-gram architectures. We explain the famous vector arithmetic and its caveat, why it yields static embeddings that cannot handle polysemy, and where it sits against the contextual embeddings of 2018.
WordNet
The lexical database of English started at Princeton in 1985 by George A. Miller’s team: 117,659 synonym sets linked by semantic relations such as hypernymy. Frozen in its official version, it lives on as a structured resource in the LLM era.
X
Y
Z
No results found.