Learning from Labeled Examples | AI Learning Guide

The key idea is that the correct answer is handed to the model along with each example. In the email example, every message carries a human-assigned tag, so when the model predicts a tag it can be compared directly against the stored one. That comparison is the error signal, and it is available for every single example, which is why supervised learning is the easiest style to reason about. Notice the asymmetry: the tag exists only while training. Once the model is in use, it has to produce the tag from the message alone, which is exactly the generalization problem from the previous chapter.

0:00 / 0:00

Auto-play

Supervised learning trains on input-output pairs where the correct output, called the label, is supplied with each example. The label is what makes the error signal direct and immediate.

A labeled dataset in practice

An email dataset where each message is stored together with a human-assigned tag, spam or not spam, is a labeled dataset. During training the model predicts a tag for a message, compares it with the stored tag, and the mismatch drives the adjustment. At use time the tag is no longer available, so the model must supply it from the message alone.

The label is only present during training. If labels are missing, expensive, or wrong, supervised learning degrades — this is the practical cost of this style.