Gradient clipping

An algorithmic operation that limits the magnitude of gradients or parameter updates during optimization.

An algorithmic operation that limits the magnitude of gradients or parameter updates during optimization.

Clipping rules can act globally, per parameter, or relative to another scale in the learning process.

Related publications