Sassha: Sharpness-aware Adaptive Second-order Optimization with Stable Hessian Approximation

Dahun Shin, Dongyeop Lee, Jinseok Chung, Namhoon Lee

Keywords: deep learning, second-order optimization, sharpness minimization

TL;DR: We introduce Sassha, a novel second-order optimization method that improves generalization by reducing solution sharpness, achieving competitive performance across diverse deep learning tasks.

Abstract:

Approximate second-order optimization methods have gained attention due to their low computational and memory overhead.
While these methods have the potential to accelerate neural network training, they often exhibit poorer generalization compared to first-order approaches. To address this limitation, we first analyze existing second-order methods through the lens of the loss landscape, demonstrating that their reduced generalization performance is somewhat attributed to the sharpness of the solutions they converge to. In response, we introduce Sassha, a novel approach designed to enhance generalization by explicitly reducing sharpness. In fact, this sharpness minimization scheme is designed to accommodate lazy and stable Hessian updates, so as to secure efficiency and robustness besides flatness. To validate its effectiveness, we conduct a wide range of deep learning experiments including standard vision and language tasks, where Sassha achieves competitive performance. Notably, Sassha demonstrates strong generalization in noisy data settings and significantly outperforms other methods in these scenarios. Additionally, we verify the robustness ofSassha through various ablation studies.

Primary Area: optimization

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 2572