Supervised vs Unsupervised vs Reinforcement: The 3 Ways a Machine Learns
Three classic ways to teach a machine, like teaching a child three different ways: labeled flashcards, an unsorted toy box, or a game with a score. What your problem gives you decides which one applies.
Supervised learning trains on thousands of labeled examples, learning input to answer — powerful, but someone has to write every label. Unsupervised learning gets raw, unlabeled data and finds patterns itself — clustering, compression, anomaly detection. Reinforcement learning has an agent act in an environment and learn from reward and penalty over time — how AlphaGo and a robot learning to walk both work. The tell is what your problem gives you: labels mean supervised, raw data means unsupervised, an agent with a goal means reinforcement.
Transcript
There are three classic ways to teach a machine — and it's exactly like teaching a child. Pick the wrong one, and you waste everything. Do you drill it with flashcards, dump a toy box and let it sort, or turn it into a game with points? That's the three types of learning.
Supervised learning is FLASHCARDS — with the answer on the back. You show it thousands of labeled examples: this photo's a cat, this email is spam. It learns the mapping from input to answer — a category is classification, a number is regression. Powerful, but somebody has to write every flashcard.
Unsupervised learning is dumping the TOY BOX on the floor with no labels. The child just sorts it — grouping by color and shape, spotting the odd piece out. No answers, only structure it finds for itself: that's clustering, compression, anomaly detection. You learn what's in your data without ever labeling it.
Reinforcement learning turns it into a GAME. The child gets no answers — just a score. Every move earns points, or loses them. By trial and error it learns a strategy that maximizes the score over time — a dog earning treats, AlphaGo, a robot learning to walk without falling.
So the tell is what you can give it. Labeled examples, to predict a label? Flashcards — supervised. Only raw data, and want structure? The toy box. An agent taking actions toward a goal, learning from feedback? The game — reinforcement. Match the teaching style to what your problem hands you.
Flashcards, a toy box, or a game. The modern twist behind LLMs is self-supervised — the text writes the flashcards: hide the next word, the answer's just what comes next. So the one thing to remember: match how you teach to what your problem hands you — labeled answers, raw data, or a reward.