Week 13: Learning from Examples

The Microsoft sequence closes by asking us to think carefully about what AI can and cannot justify. This week continues from that point with a more mechanical question: when we say a model learns from data, what is actually changing?

We will use a tiny fictional robot-distance problem because the numbers are easy to inspect. The point is not to build a useful robot controller. The point is to make supervised learning visible enough that you can defend every step.

This chapter moves through three ideas:

1. examples contain features and, in supervised learning, labels;
2. training chooses parameters or rules from the training examples;
3. a model must be tested on evidence it did not use to fit itself.

Then you will open the Classification Playground, move a decision threshold, deliberately corrupt labels, and watch the error pattern change.

By the end of the week, you should be able to explain why a model can fit its training data and still fail on new cases—and why a good-looking score is evidence about a specific test, not proof that the system is generally correct.