The Gap Between Prediction and Answer | AI Learning Guide

Think of the error as a distance with a direction. The model says 0.6, the truth is 1, so the gap is 0.4 and the model undershot. If the model had said 0.9, the gap would be only 0.1. Notice that the same email produces different errors depending on the settings, which is exactly what makes error useful: it is a per-example signal that changes as the model changes.

0:00 / 0:00

Auto-play

For one training example, the error is the gap between the model's prediction and the correct answer, and its sign shows whether the model overshot or undershot.

Two predictions, two errors

Take a spam email whose correct answer is 1. If the model outputs 0.6, the error is 0.4. If the model outputs 0.9, the error is 0.1. Same example, different settings, different error.

The sign of the error is not decoration. A positive error means the prediction was too high; a negative error means it was too low. That direction is what later tells the model which way to nudge its settings.