TBG-Driven Minimization of Noise-Resistant Adaptive Sharpness Awareness

ICLR 2026 Conference Submission17006 Authors

Keywords: Deep neural network; sharpness-awareness minimization; noise resistance; robust algorithm

Abstract: Driven by the sharpness of the loss surface effectively indicate the generalization gap, sharpness-awareness minimization (SAM) aims at flat minima within the loss landscape. However, to protect sensitive information and enhance privacy security, noise will be added to the model, which inevitably degrades the model's generalization performance. In this paper, we introduce the time-base generator (TBG) based on discrete systems and provide a boundedness theorem for discrete systems. On this basis, we propose a noise-resistant adaptive sharpness-awareness minimization method (NRASAM) , which suppresses noise through gradient decay and historical gradient integration. Furthermore, we utilized the TBG theory to adjust the algorithm parameters, resulting in the TBG-NRASAM algorithm. We provide a rigorous theoretical analysis that confirms the convergence and noise resistance of the proposed method under noisy conditions. Extensive experiments across multiple architectures and benchmarks demonstrate that our approach consistently improves generalization and stability compared to existing SAM-based methods.

Primary Area: learning theory

Submission Number: 17006