Learning from labeled examples | AI Learning Guide

The key idea is that the label does two jobs. It tells the model what the right answer is, and it produces the error signal that tells the model how to change. The cat-and-dog example shows this in action: the model guesses dog, the label says cat, and the mismatch nudges the internal parameters so that cat becomes slightly more likely next time. One example barely moves anything, but thousands of examples push the parameters into a configuration that separates the two categories. And notice what nobody did — nobody wrote down what a whisker is. The model found whatever internal structure happened to separate the labeled examples. That is why the labels themselves matter so much: a wrong label teaches a wrong lesson.

0:00 / 0:00

Auto-play

The training loop

Training is not a single pass. The model sees an example, predicts, is told the correct answer, and adjusts. The adjustment is small, so a single example barely changes anything; the change accumulates across thousands or millions of examples until the model's predictions line up with the labels. Nothing in this loop requires a human to write down what an edge or a wheel looks like — the model discovers whatever internal structure reduces its errors.

Why the label is the whole game

The label is both the target and the teacher. It defines correctness and it generates the error that drives learning. If the labels are noisy, inconsistent, or missing a class you care about, the model will faithfully learn that noise — it has no other source of truth.

A concrete case

Suppose you want a model to distinguish cats from dogs. You show it a photo and it currently guesses "dog" with high confidence, but the label says "cat." The mismatch is large, so the parameters shift to make "cat" slightly more likely for images like this one. Show it a thousand cat photos and a thousand dog photos and the same nudge happens repeatedly, in different directions, until the model separates the two. Notice that no one told it what a whisker or a snout is — it found whatever internal features happened to separate the labeled examples.

A model can only learn categories that appear in its training labels. If every cat photo in the dataset is a tabby, the model may never learn that black cats are cats — not because it is broken, but because it was never taught.