Supervised, Unsupervised & Reinforcement

Three ways to learn

All of machine learning falls into three paradigms, distinguished by what feedback the learner gets.

1. Supervised learning: learning with an answer key

You provide inputs and correct outputs; the model learns the mapping. This is ~80% of ML in industry. Two flavors:

• Regression — predict a number: house price, temperature, delivery time.
• Classification — predict a category: spam/not-spam, cat/dog/horse, tumor benign/malignant.

The mental test: "could a human annotator write down the right answer for each example?" If yes, it's supervised.

2. Unsupervised learning: finding structure without answers

Only inputs, no labels. The model hunts for hidden structure:

• Clustering — group similar things: customer segments, topic groups in news.
• Dimensionality reduction — compress features while keeping information (PCA from the math path).
• Anomaly detection — flag the weird: fraud, failing machines, network intrusions.

Harder to evaluate (there's no "right answer"), but labels are expensive and unlabeled data is everywhere — most of the world's data is unlabeled.

3. Reinforcement learning: learning by trial and error

An agent acts in an environment, receives rewards, and learns a strategy (policy) that maximizes long-term reward. No one shows the correct action; the agent discovers it by experimenting.

$\text{state} \xrightarrow{\;\text{action}\;} \text{new state},\; \text{reward}$

This is how AlphaGo beat world champions and how robots learn to walk. The catch: RL is data-hungry and finicky — you'll study it properly in the AI path.

The modern twist: self-supervised learning

Where do ChatGPT-style models fit? They use a clever trick: take unlabeled text and manufacture labels from it — hide the next word and train the model to predict it. The labels are free, the scale is the whole internet. This self-supervised learning blurs the supervised/unsupervised line and powers the current AI boom.

Matching problems to paradigms

Key takeaways

• Supervised = learn from answer keys; unsupervised = find structure; reinforcement = learn from rewards.
• Ask "what feedback exists?" to classify any ML problem.
• Self-supervised learning manufactures labels from raw data — the engine of modern language models.