Types of Learning

Types of Learning

1. Three Paradigms

Type	Data	Goal	Example
Supervised	$(x, y)$ labelled	learn $x → y$	spam, price prediction
Unsupervised	$x$ only	find structure	clustering, PCA
Reinforcement	state, action, reward	learn policy $π$	game play, robot

2. Supervised Split

• Classification: $y$ discrete (spam/ham, digit 0-9). Metric: accuracy, F1.
• Regression: $y$ continuous (price, temp). Metric: squared error.

3. Unsupervised Split

• Clustering: group similar ($k$-means).
• Density / anomaly: flag rare points.
• Dimensionality reduction: $R^{d} → R^{k}$, $k < d$ (PCA).

4. Other Settings

Semi-supervised (few labels + many unlabelled), self-supervised (mask + predict), transfer (reuse pretrained), online (streaming update).

5. Exam Points

• Compare supervised vs unsupervised vs RL with examples — 5-7 marks, very frequent.
• Classification vs regression — 3 marks, frequent.
• One example each of clustering, reduction, RL — 3 marks.