May, 2023

锐度感知最小化导致低秩特征

TL;DRSharpness-aware minimization (SAM) method can reduce feature ranks in various types of neural networks, and the phenomenon is observed in a simple two-layer network. A significant number of activations gets pruned by SAM, which contributes directly to this rank reduction. The observed low-rank effect can also occur in deep networks, although the overall mechanism can be more complex.