Skip to main content
All terms
Architectures

Residual Connection

A shortcut that adds a layer's input to its output, helping deep networks train stably.

Definition

A residual connection, also called a skip connection, adds a layer's input directly to its output so that the layer only has to learn a change to the input. This gives gradients a clean path back through the network, which makes it possible to train models hundreds of layers deep. Introduced in ResNets, residual connections wrap both sublayers of every Transformer block.