Conclusion

This chapter built the first small neural network from familiar parts.

An MLP composes affine maps and nonlinear activations. Hidden layers produce intermediate activations. Depth controls how many learned stages the model has. Width controls how many units a layer computes. Hidden representations stand between raw input and final prediction.

The chapter intentionally stopped before losses and training. We can now trace an MLP forward pass. The next step is to ask how to judge the prediction.

That is the role of loss functions.