δ-SAM: Sharpness-Aware Minimization with Dynamic Reweighting
share
Deep neural networks are often overparameterized and may not easily achieve model generalization. Adversarial training has shown effectiveness in improving generalization by regularizing the change of loss on top of adversarially chosen perturbations. The recently proposed sharpness-aware minimization (SAM) algorithm adopts adversarial weight perturbation, encouraging the model to converging to a flat minima. Unfortunately, due to increased computational cost, adversarial weight perturbation can only be efficiently approximated per-batch instead of per-instance, leading to degraded performance. In this paper, we propose that dynamically reweighted perturbation within each batch, where unguarded instances are up-weighted, can serve as a better approximation to per-instance perturbation. We propose sharpness-aware minimization with dynamic reweighting (δ-SAM), which realizes the idea with efficient guardedness estimation. Experiments on the GLUE benchmark demonstrate the effectiveness of δ-SAM.
READ FULL TEXT