A residual connection lets a neural-network layer add a transformation to an existing representation instead of replacing it completely. A gated residual connection adds a learned control that regulates the size or content of that update. The model can preserve an earlier signal when a new transformation would be unhelpful.
Gating can improve information flow through deep networks and reduce repeated overwriting across layers. It also gives the model another mechanism to decide when to retain or modify context. The exact gate design varies by architecture, and its benefits must be measured against the added parameters and computation.
ELI5
A gated residual connection lets a neural-network layer decide how much new information should change the representation it already has. A learned gate can preserve the earlier signal when the new calculation is not useful.
For example, a deep model can carry an important token feature through several layers and open the gate only when a later transformation adds value. The extra control can improve information flow, but it also adds parameters and computation that need testing.
