A shortcut that adds a layer's input directly onto its output, so a deep stack of layers only has to learn a correction on top of an identity path instead of the whole transformation from scratch, which keeps gradients flowing during training.
See it explained in full