Groups
Category
Sharpness-Aware Minimization (SAM) trains models to perform well even when their weights are slightly perturbed, seeking flatter minima that generalize better.
Spectral analysis studies the distribution of eigenvalues and singular values of neural network weight matrices during training.