~/learn/Machine Learning · Foundations & generalization
Learning paradigms
Supervised, unsupervised, self-supervised and reinforcement learning — what each one learns from.
why this matters
Interviews often open with 'how would you frame this problem?'. Naming the right paradigm and label source shows you think before you model.
A model learns a function from data. The paradigm is defined by what signal it learns from. Supervised learning uses labelled pairs (x, y): predict churn from usage, price from features. Regression predicts a number; classification predicts a category.
Unsupervised learning has no labels — it finds structure: clusters of customers, low-dimensional representations, anomalies. Self-supervised learning invents labels from the data itself (predict the next word, the masked pixel) and is how large language models are pre-trained. Reinforcement learning learns from rewards after actions, not from correct answers.
Pick by the signal
labelled (x, y) → supervised → predict y only x → unsupervised → find structure x with hidden part → self-supervised → predict the hidden part actions + rewards → reinforcement → maximise reward
Interview trap
'We have no labels' rarely means 'use clustering'. Often the right answer is to collect a small labelled set, use weak labels (rules, heuristics), or reframe the task.
key takeaways
- Supervised learns from labels; unsupervised finds structure.
- Self-supervised creates labels from the data itself.
- Reinforcement learning learns from rewards.
check your understanding
A team pre-trains a model by hiding words in sentences and predicting them. Which paradigm is this?