SADT: Combining Sharpness-Aware Minimization with Self-Distillation for Improved Model Generalization

TL;DR

This paper introduces SADT, a novel training strategy that combines sharpness-aware minimization and self-distillation, leading to improved model convergence, performance, and generalizability across diverse neural architectures and datasets.

Contribution

The paper proposes SADT, a new training method that effectively integrates sharpness-aware minimization with self-distillation, enhancing deep neural network training.

Findings

SADT outperforms existing strategies in convergence speed.

SADT improves test-time performance across multiple datasets.

SADT enhances model generalizability over various architectures.

Abstract

Methods for improving deep neural network training times and model generalizability consist of various data augmentation, regularization, and optimization approaches, which tend to be sensitive to hyperparameter settings and make reproducibility more challenging. This work jointly considers two recent training strategies that address model generalizability: sharpness-aware minimization, and self-distillation, and proposes the novel training strategy of Sharpness-Aware Distilled Teachers (SADT). The experimental section of this work shows that SADT consistently outperforms previously published training strategies in model convergence time, test-time performance, and model generalizability over various neural architectures, datasets, and hyperparameter settings.