Supervised learning

We like to think of supervised learning as teaching a friend to spot cats in photos. We hand her a pile of pictures with sticky notes: “cat” or “not cat.” She studies. Then we quiz her with new photos. If she gets it right, we nod. If not, we adjust.

Features

In this setup, features are the clues. For a photo, they might be whiskers, pointy ears, or the shape of the tail. In code, features are just numbers—pixel values, word counts, temperature readings. They don’t look exciting, but they’re the bread and butter of training. Without them, she’s guessing in the dark.

Labels

Labels are the answers we give away. Each picture has one: “cat” or “not cat.” In real projects, labels could be “spam” versus “ham” emails, or “approved” versus “denied” loan applications. The catch is we need a lot of them, and labeling takes time. It’s the grunt work no one brags about, but without it she doesn’t learn.

Training

Training is the repetitive grind. We feed her thousands of feature–label pairs, and she tweaks her internal knobs to minimize mistakes. We call it optimization, but it feels more like practice: miss, adjust, try again. Eventually, her error rate drops. She starts spotting cats reliably. We don’t expect perfection; we settle for good enough that she won’t embarrass us in production.

Our takeaway

Supervised learning is straightforward. Give her examples with answers, let her practice, and check how she does on new stuff. The cleverness isn’t in the math—though that helps—it’s in setting up features and labels that actually match the problem we care about. Otherwise, we end up training her to chase the wrong mouse.