Imagine a model that predicts a house's selling price from one input, its size. We train it on nine past sales. Each dot on the plot is one example: the input along the bottom, the true output up the side. The dots follow a trend, but with scatter, because real data always carries some noise: luck, a typo, a quirky buyer.
The Trainer adjusts the Model until its predictions sit close to the dots. How close is measured by the training error, the average gap between each dot and the model's line. Pushing that number down looks like exactly what we want. This lesson shows why it is not enough.
💡Good to know
Training error says how well the model fits examples it has seen, not how it will do on new ones.