Bilateral Sharpness-Aware Minimization for Flatter Minima
Bilateral Sharpness-Aware Minimization (BSAM) improves model generalization by combining Max-Sharpness (MaxS) and Min-Sharpness (MinS). Standard SAM fails due to a 'Flatness Indicator Problem' because it only checks flatness in one direction. BSAM uses both measures to find flatter minima, leading to superior performance across various tasks like classification and quantization.