Supervised, Unsupervised & Reinforcement Learning

Every ML problem falls into one of three buckets. Knowing which one you have decides everything that follows.

1. Supervised learning — learning from labelled examples

You have inputs and the correct answers. The model learns the mapping. This is 90% of applied ML.

• Classification — predict a category. Spam/not-spam, disease/healthy, which digit is this?
• Regression — predict a number. House price, tomorrow's temperature, CGPA.

2. Unsupervised learning — finding structure with no labels

You have data but no answers. The model finds hidden patterns.

• Clustering — group similar things. Customer segments, news topics.
• Dimensionality reduction — compress features while keeping meaning (PCA, embeddings).

3. Reinforcement learning — learning by trial and reward

An agent takes actions in an environment and gets rewards/penalties. It learns a strategy that maximises reward over time. Powers game-playing AI (AlphaGo), robotics, and the "RLHF" step that made ChatGPT helpful.

Problem: "Predict if this transaction is fraud"       -> Supervised (classification)
Problem: "Group these 1M users into segments"          -> Unsupervised (clustering)
Problem: "Teach a bot to win at chess"                 -> Reinforcement
Problem: "Estimate a used car's resale price"          -> Supervised (regression)