AdamW
AdamW: An Adam variant separating weight decay from the adaptive update based on gradient moments. Decoupling prevents the L2 penalty from being scaled like the gradient. Learning rate and decay coefficient remain distinct hyperparameters. This perspective connects the definition with practical behaviour in artificial intelligence systems.
In italiano: AdamWLegge IV · Loss, gradiente e ottimizzatoriDecoupling prevents the L2 penalty from being scaled like the gradient. Learning rate and decay coefficient remain distinct hyperparameters.