Residual connection
Also: skip connection, shortcut connection
Instead of learning a mapping H(x) directly, a residual block learns F(x) = H(x) − x and outputs F(x) + x. The identity path gives gradients a direct route backward, which is what made networks with 100+ layers trainable (ResNet) and is used in every Transformer block.
y = F(x) + x
Example
If the best a block can do is pass its input through, it only needs F(x) = 0. Its gradient ∂y/∂x = ∂F/∂x + 1 never vanishes, which let ResNet-152 train with 152 layers.