Learning Paradigms
The different ways a machine can learn: from labeled answers, from raw unlabeled data, from data that labels itself, or from rewards.
There are different ways to learn to cook. A teacher can show you the right dish each time, you can sort ingredients into piles on your own, or you can just try things and see whether people enjoy the meal.
Where it fits
Not shown: 5 more kinds and parts. See the full list below.
+All relationships as a list
Overview
A paradigm is not a model. It describes what kind of feedback the learner gets, and the same model can be used under different paradigms. A neural network can be trained with labels, without them, or with rewards.
- Supervised learning: every example comes with the right answer. Predict prices, classify emails.
- Unsupervised learning: no answers at all. Find groups, compress, detect what is unusual.
- Self-supervised learning: the data supplies its own answers, for example “predict the next word”. This is how large language models are pretrained.
- Transfer learning: start from a model trained on one task and adapt it to another, the default way to build models today.
- Reinforcement learning: an agent acts, receives rewards, and learns a strategy. It powers game-playing systems and parts of RLHF.
Everyday use cases
- Case 01
Choosing an approach
Do you have labels? A reward? Only raw data? The answer tells you which family of methods to reach for.