The property of residual connections that allows gradients to flow directly through the identity path during backpropagation, bypassing layer transformations. This enables stable training of very deep networks.

See also: residual_connections, vanishing_gradients, attention_residuals