Residual connection

Also: skip connection

Adding a sublayer's input to its output, so the sublayer only has to learn a correction rather than the whole mapping. Without it, gradients through a deep stack shrink to nothing and training stalls.