6 / 6 Activations & Optimization

What objective does SAM approximately optimize?

Sharpness-Aware Minimization modifies ordinary empirical-risk minimization by approximately minimizing which objective?

A

The average loss of independently initialized models lying on a low-loss path

B

The expected loss after adding isotropic Gaussian noise only to the input features

C

The maximum training loss within a bounded neighborhood of the current parameters

D

The trace of the exact Hessian subject to zero training loss