[edit]
On the Maximum Hessian Eigenvalue and Generalization
Proceedings on "I Can't Believe It's Not Better! - Understanding Deep Learning Through Empirical Falsification" at NeurIPS 2022 Workshops, PMLR 187:51-65, 2023.
Abstract
The mechanisms by which certain training interventions, such as increasing learning rates and applying batch normalization, improve the generalization of deep networks remains a mystery. Prior works have speculated that "flatter" solutions generalize better than "sharper" solutions to unseen data, motivating several metrics for measuring flatness (particularly $\lambda_{\rm max}$ , the largest eigenvalue of the Hessian of the loss); and algorithms, such as Sharpness-Aware Minimization (SAM), that directly optimize for flatness. Other works question the link between $\lambda_{\rm max}$ and generalization. In this paper, we present findings that call $\lambda_{\rm max}$’s influence on generalization further into question. We show that: (1) while larger learning rates reduce $\lambda_{\rm max}$ for all batch sizes, generalization benefits sometimes vanish at larger batch sizes; (2) by scaling batch size and learning rate simultaneously, we can change $\lambda_{\rm max}$ without affecting generalization; (3) while SAM produces smaller $\lambda_{\rm max}$ for all batch sizes, generalization benefits (also) vanish with larger batch sizes; (4) for dropout, excessively high dropout probabilities can degrade generalization, even as they promote smaller $\lambda_{\rm max}$ ; and (5) while batch-normalization does not consistently produce smaller $\lambda_{\rm max}$ , it nevertheless confers generalization benefits. While our experiments affirm the generalization benefits of large learning rates and SAM for minibatch SGD, the GD-SGD discrepancy demonstrates limits to $\lambda_{\rm max}$’s ability to explain generalization in neural networks.