Three ways to learn

"Supervised," "unsupervised," "reinforcement" — three intimidating words that get thrown around constantly, usually without anyone explaining what actually separates them. Here's the good news: the difference comes down to one simple question — what kind of feedback does the learner get? Get that, and the whole map clicks into place.

Everything in machine learning is a system trying to improve at a task. What makes the three paradigms different is what tells it whether it's doing well.

Scroll to see the one thing that sets them apart.

The feedback is the whole story

Three everyday analogies capture it:

• Supervised is studying with flashcards: each card has a question and the right answer. You practice until your guesses match the answers, then face new cards. The feedback is a label — the correct output for every example.
• Unsupervised is tidying a junk drawer nobody has sorted. No one tells you the categories; you just notice that these things go together and those don't. The feedback is… none. You work from the structure of the stuff itself.
• Reinforcement is learning a video game with no manual. You press buttons, and all you get back is the score going up or down. The feedback is a reward — delayed, sparse, and about outcomes rather than individual right answers.

Why the distinction matters

Picking the wrong paradigm dooms a project before it starts. No labels? Supervised learning is off the table. Need a system that acts and improves over time, not just predicts once? That's reinforcement's turf. Just want to understand a big messy dataset? Unsupervised. Knowing which is which is the first real skill in machine learning.

Over the next three chapters we'll take each one in turn — starting with the most common and most intuitive: supervised learning, where the machine learns from worked answers.