Normalization Layers Are All That Sharpness-Aware Minimization Needs

Maximilian Mueller, Tiffany Joyce Vlaar, David Rolnick, Matthias Hein

Keywords: sharpness-aware minimization, flatness, generalization, normalization layers

TL;DR: We show that only applying SAM to the normalization layers of the network enhances performance for ResNet and Vision Transformer architectures compared to applying SAM to all parameters.

Abstract:

Sharpness-aware minimization (SAM) was proposed to reduce sharpness of minima and has been shown to enhance generalization performance in various settings. In this work we show that perturbing only the affine normalization parameters (typically comprising 0.1% of the total parameters) in the adversarial step of SAM can outperform perturbing all of the parameters. This finding generalizes
to different SAM variants and both ResNet (Batch Normalization) and Vision Transformer (Layer Normalization) architectures. We consider alternative sparse perturbation approaches and find that these do not achieve similar performance enhancement at such extreme sparsity levels, showing that this behaviour is unique to the normalization layers. Although our findings reaffirm the effectiveness
of SAM in improving generalization performance, they cast doubt on whether this is solely caused by reduced sharpness.

Submission Number: 14433