Asynchronous Sharpness-Aware Minimization For Fast and Accurate Deep Learning

Name: Asynchronous Sharpness-Aware Minimization For Fast and Accurate Deep Learning
Uploaded: 2026-10-15T12:34:13.301552+00:00
Duration: 276 s
Description: The method addresses the high computational cost of Sharpness-Aware Minimization (SAM) by introducing asynchronous model perturbation and system-aware batch size adjustment. This allows training to fully utilize heterogeneous resources like CPUs and GPUs, effectively hiding the perturbation time while maintaining or improving generalization performance.

The method addresses the high computational cost of Sharpness-Aware Minimization (SAM) by introducing asynchronous model perturbation and system-aware batch size adjustment. This allows training to fully utilize heterogeneous resources like CPUs and GPUs, effectively hiding the perturbation time while maintaining or improving generalization performance.

Open the full episode