Conclusion

This chapter connected predictions to learning signals.

Prediction error tells direction. Mean squared error gives regression models a number that grows with squared mistakes. Binary cross-entropy and multiclass cross-entropy reward assigning high probability to the correct class. Logits are raw scores; probabilities are normalized values. Softmax turns multiclass logits into probabilities, and stable softmax avoids unnecessary numerical trouble.

The next chapter asks how computation is recorded so gradients can be computed.