Introduction

Deep networks need paths for information to move.

If every layer must transform everything, useful signals can be distorted or weakened before they reach later layers. Gradients can also struggle to move backward through many transformations.

Skip, residual, and gated connections add routes around or through blocks.

The idea is simple: a layer does not always need to replace its input. It can add to it, pass it forward, concatenate it, or learn how much to keep.

This chapter explains those connection patterns before larger architectures hide them behind names.