Machine learning families
The four learning paradigms plus the architecture they run on. Label availability is the sorting key: supervised uses labelled pairs, unsupervised finds structure in unlabelled data, semi-supervised blends both, and reinforcement learning optimises for reward.
Machine learning is the AI subset where systems learn patterns from data to improve at tasks without being explicitly programmed rule by rule. The output of that learning is the Machine learning model - the trained artifact of learned parameters and logic. The four paradigms differ on one axis above all: how much of the data is labelled.
| Paradigm | Data | What it produces |
|---|---|---|
| Supervised learning | Labelled input-output pairs | Classification and regression |
| Unsupervised learning | Unlabelled data | Clustering and association (hidden structure) |
| Semi-supervised learning | Small labelled set + large unlabelled set | Where the course files LLMs |
| Reinforcement learning | Reward and penalty signals | Policy learned by trial and error |
- Reinforcement learning with human feedback (RLHF) is RL whose reward signal is human preference ratings - the alignment step that makes outputs helpful and appropriate.
- Deep learning is ML built on many-layered Neural networks that learn hierarchical representations - the engine behind generative AI.
When a scenario asks which paradigm fits, ask first: are the examples labelled? Labelled → supervised; none → unsupervised; a few → semi-supervised; rewards instead of labels → reinforcement.