.png){kind=icon}
.png){kind=icon}
Return
A Two-Timescale Neurodynamic Approach to Sharpness-Aware Minimization in Deep Learning
.png){kind=icon}
PRE
AI
DOI：10.1109/JAS.2026.126020.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Dear Editor, This letter presents a two-timescale neurodynamic algorithm for sharpness-aware minimization in deep learning. Deep learning achieves remarkable success in areas such as computer vision, natural language processing, robotics and control. In deep learning, it is essential to boost their generalization power [1]. Existing deep learning strategies for improving the generalization power include regularization, data augmentation, etc. [2]. Sharpness-aware minimization (SAM) is a deep learning strategy that seeks flat regions in the loss landscape to improve the generalization power of deep neural networks (DNNs). Nevertheless, the existing SAM algorithms with tightly coupled updating of DNN parameters and perturbations often lead to instability and hyperparameter sensitivity. The proposed algorithm addresses SAM from a timescale perspective, aiming to decouple perturbation and parameter dynamics in short and long timescales.
Journal
I
IF:
0
Papers:
116
Citations:
0