Conclusion

This chapter made the first model family concrete.

A linear model computes scores with weighted sums and biases. Those scores can be used for binary or multiclass decisions. In two dimensions, a binary linear classifier separates space with a line. In higher dimensions, the same idea becomes a hyperplane.

The strength of linear models is also their limit. They are simple and inspectable, but one linear boundary cannot represent every useful pattern.

The next chapter adds nonlinear activations, which let networks build richer functions from linear pieces.