The AI atlas: what belongs to AI, and what it connects to
An expanded concept map of all the main layers of AI and its neighbouring disciplines.
A note first: there is no single, universally accepted way to classify AI. “Machine learning” describes a method, “robotics” a field of action, “healthcare” an area of use, and “ethics” a question of responsibility. So they cannot all fairly be placed on one ladder.
A. The wider scientific foundations
| Parent field or related discipline | What it brings to AI | Where to start |
|---|---|---|
| Computer science and software engineering | Algorithms, systems, programming, security. | Algorithm or model? |
| Mathematics | Logic, linear algebra, optimisation, probability. | Mathematics and statistics |
| Statistics and data science | Measuring uncertainty, testing, drawing conclusions. | Causality |
| Cognitive science and psychology | Models of learning, memory and problem-solving. | Mind and brain |
| Neuroscience | Biological inspiration and an understanding of nervous systems. | Neural networks |
| Linguistics | Meaning, grammar, how language is used, communication. | Natural language processing |
| Philosophy and logic | Knowledge, reasoning, meaning, consciousness, ethics. | Can AI be conscious? |
| Robotics and control theory | Sensors, movement, planning and control. | Robotics |
| Economics, law and social sciences | Effects, incentives, rights and accountability. | AI and society |
B. The main methods and technical sub-fields
| Sub-field | Start with |
|---|---|
| Symbolic AI, logic, knowledge representation | Rules and knowledge |
| Search, planning, optimisation | Choosing a route or an action |
| Machine learning (ML) | Learning from examples |
| Supervised, unsupervised, self-supervised | Ways of learning |
| Deep learning and neural networks | Layers in models |
| Reinforcement learning (RL) | Learning from rewards |
| Causal AI | Causes, not just patterns |
| Generative models | Creating content |
| Language models (LLMs) and NLP | Language |
| Computer vision | Images and video |
| Speech and audio technologies | Speech and sound |
| Multimodal models | Different kinds of input |
| Agents, multi-agent systems, tool use | From answering to acting |
| Robotics and embodied AI | Acting in the physical world |
C. What is not a sub-field in the same sense?
- Healthcare, education, art, media, transport: areas of use; they can draw on many different AI techniques.
- Security, privacy, ethics, law: requirements and questions that cut across every system.
- GPUs, data centres, APIs: infrastructure and ways of connecting things, not a separate “type of intelligence”.
- AGI: a term for a hypothetical level of general ability, not an established technical branch.
D. Smaller specialised fields — where they fit
These approaches partly overlap. Each one starts with a simple question.
- {L0} — Bayes and estimating uncertainty.
- {L0} — classic machine-learning methods.
- {L0} — knowledge graphs and ontologies.
- {L0} — models of relationships in networks.
- {L0} — trial, selection and improvement.
- {L0} — working together without pooling all the data in one place.
- {L0} — recommendations, ranking and anomalies.
- {L0} — forecasts and their errors.
- {L0} — understanding predictions.
- {L0} — simulations and the possible consequences of actions.
- {L0} — distillation and quantisation.
Behaviour-based robotics, geometric learning, neuro-symbolic systems and other smaller research directions are linked to these main branches. They cannot be placed in one tree with neat, separate boxes.
These fields matter; for now we introduce them on the atlas, and more detailed explanations will follow based on reader interest and checked sources.
Remember: AI can be classified in at least four ways: by method, by input data, by task and by social context.
Last reviewed 1 November 2026. An explainer, not professional, legal or medical advice. For rules and figures that change, check the official version.