A Collaborative Neurodynamic Approach to Sharpness-Aware Minimization for Improving Generalization of Deep Learning

Boosting generalization in deep neural networks (DNNs) is a central challenge in deep learning, especially for highly overparameterized neural networks. Sharpness-aware minimization (SAM) addresses this issue by guiding learning toward flat regions of the loss landscape. However, the original SAM algorithm relies on a locally linearized loss around an anchor point to determine the maximum sharpness, which compromises generalization performance and makes it sensitive to the choice of the anchor point. To address this issue, we propose a collaborative neurodynamic approach to SAM, called CNSAM. To diversify exploration, CNSAM leverages multiple agents to search for flat regions of the loss landscape, leading to improved generalization. Extensive experiments on image classification and natural language processing (NLP) demonstrate that CNSAM consistently achieves a lower test loss and a higher accuracy than SAM and other related learning algorithms.