The examples decide what gets learned | AI Learning Guide

The key thing to hold onto is that the model never sees a rule. It sees examples with answers attached, and it works backward to a rule. That means the examples are doing the teaching, so their flaws become the model's flaws. Walk through the three properties. If only a few people carry the disease label, the model may latch onto something incidental about those particular people. If one label hugely outnumbers the other, always guessing the common answer looks accurate but tells you nothing. And if some people in the healthy group were actually sick but undiagnosed, the answer key itself is wrong — a bigger model will just learn the error more confidently.

0:00 / 0:00

Auto-play

Training data is the collection of labeled examples a learning system is shown. The model does not receive the rule; it receives examples and their correct answers, and it works backward from those to a rule of its own.

What sets the ceiling

Size — a pattern supported by only a few examples cannot be told apart from coincidence
Balance — if one label dominates, the model can score well by always guessing it
Label accuracy — a wrong answer key teaches the model to be confidently wrong, and more data does not repair it

A more powerful model does not compensate for poor training data. It will fit the data it was given, including its errors, more thoroughly.