The world of AI.

Supervised learning — learns from labeled examples to predict outcomes on new data. Includes linear and logistic regression, decision trees, random forests, support vector machines, naive Bayes, k-nearest neighbours, and gradient boosting (e.g. XGBoost).

Unsupervised learning — finds structure in unlabeled data. Includes k-means and hierarchical clustering, DBSCAN, principal component analysis (PCA), t-SNE, and association-rule mining (Apriori).

Reinforcement learning — learns through trial, error, and reward. Includes Q-learning, deep Q-networks, policy-gradient methods (PPO), actor-critic methods, and Monte Carlo tree search.

Semi- and self-supervised learning — combines a small labeled set with a large unlabeled one. Includes co-training, self-training, and contrastive learning.

Deep learning / neural networks — layered networks that learn hierarchical representations. Includes CNNs (vision), RNNs and LSTMs (sequences), transformers (language), autoencoders, and generative adversarial networks (GANs).

Evolutionary and bio-inspired methods — optimisation inspired by natural processes, including genetic algorithms and particle swarm optimisation.

Time series & forecasting — modelling data that changes over time, from classical methods like ARIMA to modern sequence-based forecasting models.

Language understanding — extracting meaning from text: named entity recognition, information extraction, sentiment analysis, and intent classification.

Speech — converting between speech and text, including speech recognition (ASR), text-to-speech (TTS), and speaker identification.

Machine translation — automatically translating between languages, from early statistical methods to modern transformer-based neural translation.

Large language models — models trained on massive text corpora for generation, reasoning, and dialogue.

Text generation and summarisation — producing coherent new text or condensing longer documents into their key points.

Dialogue systems — conversational agents that manage multi-turn interaction and context.

Image classification — identifying what an image contains, the basis of most object recognition systems.

Object detection & segmentation — locating and outlining individual objects within a scene.

3D reconstruction — building 3D models from 2D images using motion, shading, and contour analysis.

Facial recognition & biometrics — identifying or verifying individuals from facial or biometric data.

Optical character recognition (OCR) — reading printed or handwritten text from images.

Video understanding — recognising actions and tracking objects across video frames over time.

Perception — sensing and interpreting the environment through cameras, LiDAR, and other sensors.

Localisation & mapping (SLAM) — determining a robot's position while simultaneously building a map of its surroundings.

Motion planning & control — planning collision-free paths and executing precise, coordinated movement.

Actuation — converting decisions into physical action via motors and mechanical effectors.

Human-robot interaction — designing safe, intuitive collaboration between robots and people.

Probabilistic reasoning — reasoning under uncertainty using probability theory rather than strict logic.

Bayesian networks — graphical models representing probabilistic relationships between variables.

Temporal reasoning — reasoning about how facts and states change over time.

Uncertainty quantification — measuring how confident a model should be in its own predictions.

Fuzzy logic — reasoning with degrees of truth rather than a strict true/false.

Case-based reasoning — solving new problems by adapting solutions to similar past problems.

Logic-based systems — formal rules for deriving conclusions from known facts.

Knowledge engineering — structuring domain expertise into a form a system can use.

Automated planning & scheduling — generating a sequence of actions that leads from a start state to a goal.

Real-world planning — planning under real constraints like time, cost, and limited resources.

Game theory — modelling strategic decision-making between multiple rational agents.

Search algorithms — systematically exploring possibilities to find a solution, e.g. breadth-first, depth-first, and A* search.

Heuristic search — using rules of thumb to search more efficiently than exploring every possibility.

Constraint satisfaction — solving problems defined by a set of constraints that must all hold.

Adversarial search — searching in competitive settings against an opponent, such as minimax and game trees.

Optimisation — finding the best solution among many feasible options.

What it is — systems with multiple autonomous agents that interact with each other and their environment, learning and adapting their behaviour over time to achieve a common goal.

Coordination & communication — how agents share information, negotiate, and divide labour.

Emergent behaviour — complex group behaviour arising from simple individual agent rules.

Distributed problem solving — splitting a task across many cooperating agents rather than one central solver.

Large language models — generating text, code, and reasoning from massive pretraining on language.

Image & video generation — diffusion models and GANs producing novel visual content from a prompt.

Audio & music generation — models composing speech, sound effects, or music.

Retrieval-augmented generation (RAG) — grounding a model's output in external, up-to-date knowledge sources.

Ontologies & semantic networks — structured representations of concepts and the relationships between them.

Rule-based expert systems — encoding human expert knowledge as if-then rules to automate specialist decisions.

Knowledge graphs — connecting entities and facts into a queryable graph structure.

Neuro-symbolic AI — combining neural networks' pattern recognition with symbolic reasoning's logical rigor.

Genetic algorithms — evolving candidate solutions through selection, crossover, and mutation across generations.

Swarm intelligence — decentralised systems inspired by ant colonies, bird flocks, and other collective animal behaviour.

Particle swarm optimisation — optimising by moving a population of candidate solutions through a search space together.

Fairness & bias mitigation — identifying and reducing unfair or discriminatory model behaviour.

Adversarial robustness — defending models against inputs deliberately crafted to fool them.

Explainable AI (XAI) — making a model's decisions interpretable to the humans relying on them.

AI alignment & safety — ensuring AI systems reliably pursue the goals intended for them.

Governance & regulation — the policy and legal frameworks shaping responsible AI use.

Emotion recognition — detecting human emotion from voice, facial expression, or text.

Sentiment & opinion mining — extracting attitudes and opinions expressed in text.

Human-centred AI design — building systems that respond appropriately to a person's emotional state.

Collaborative filtering — recommending based on patterns across many users' preferences, not just one person's history.

Content-based filtering — recommending items similar in attributes to what a user already likes.

Matrix factorisation — decomposing user-item interactions to uncover latent preferences, the approach that won the Netflix Prize.

Hybrid & deep learning-based systems — combining multiple signals, including learned embeddings, for large-scale personalisation.

Hyperparameter optimisation — automatically searching for the best model settings instead of tuning them by hand.

Neural architecture search (NAS) — automatically designing neural network architectures rather than hand-crafting them.

Meta-learning — training models that learn how to learn, adapting quickly to new tasks with very little data.

Few-shot & transfer learning — reusing knowledge from one task, or a pretrained model, to solve a new one with minimal additional data.

Federated learning — training a shared model across many devices or servers without ever moving the raw data off them.

Differential privacy — adding calibrated noise so a model can't leak information about any single individual in its training data.

Secure multi-party computation — multiple parties jointly computing a result without revealing their private inputs to each other.

Homomorphic encryption — computing directly on encrypted data without ever needing to decrypt it.