Bilateral Sharpness-Aware Minimization for Flatter Minima.

Area of Science:

• Deep Learning
• Optimization
• Computer Vision

Background:

• Sharpness-aware minimization (SAM) improves deep neural network (DNN) generalization by minimizing max-sharpness (MaxS).
• SAM's MaxS metric suffers from the flatness indicator problem (FIP), inadequately assessing solution region flatness and leading to high Hessian eigenvalues.
• This indicates insufficient flatness, hindering optimal generalization and robustness.

Purpose of the Study:

• To address the flatness indicator problem in SAM.
• To develop a novel flatness indicator that better reflects the solution region's geometry.
• To enhance DNN generalization and robustness through improved optimization.

Main Methods:

• Introduced min-sharpness (MinS) as a new flatness indicator, measuring the difference between training and minimum neighborhood loss.
• Proposed bilateral SAM (BSAM), integrating MaxS and MinS for a more comprehensive flatness assessment.
• Conducted theoretical analysis to demonstrate BSAM's convergence properties.

Main Results:

• BSAM identifies flatter minima compared to vanilla SAM.
• Extensive experiments across diverse tasks (classification, transfer learning, pose estimation, semantic segmentation, quantization) show BSAM's superiority.
• BSAM demonstrates improved generalization performance and robustness over SAM.

Conclusions:

• The proposed min-sharpness metric effectively addresses SAM's flatness indicator problem.
• Bilateral SAM (BSAM) offers a more robust and effective optimization strategy for deep neural networks.
• BSAM achieves superior generalization and robustness across a wide range of computer vision tasks.