Why ResNet uses residual blocks AI

How a learned transformation and a shortcut combine within the image network.

ResNet-18 organizes image transformations into residual blocks. A block adds its learned transformation to a shortcut from its input. Conceptually, the output is a refinement of what entered the block:

$$y=\operatorname{ReLU}(F(x)+S(x)).$$

$F$ denotes the convolutional transformation; $S$ is the shortcut. When shapes match, the shortcut can be the identity. When resolution or channel count changes, a learned projection can make the shapes compatible. The original ResNet paper motivates residual learning as a way to make deep networks easier to optimize.

In this inspector, layer1 through layer4 are stages containing residual blocks. The saved stage maps are outputs after the last block’s addition and ReLU. Intermediate convolution outputs and shortcut responses were not separately saved. A stage map therefore cannot show how much of its value arrived by each route.

Depth allows repeated transformations of increasingly broad image context. It does not guarantee that each layer corresponds to a named visual concept.

Explore the neural-network demo.