GD doesn't make the cut: Three ways that non-differentiability affects neural network training

This paper investigates the distinctions between gradient methods applied to non-differentiable functions (NGDMs) and classical gradient descents (GDs) designed for differentiable functions. First, we demonstrate significant differences in the convergence properties of NGDMs compared to GDs, challenging the applicability of the extensive neural network convergence literature based on $L-smoothness$ to non-smooth neural networks. Next, we demonstrate the paradoxical nature of NGDM solutions for $L_{1}$-regularized problems, showing that increasing the regularization penalty leads to an increase in the $L_{1}$ norm of optimal solutions in NGDMs. Consequently, we show that widely adopted $L_{1}$ penalization-based techniques for network pruning do not yield expected results. Finally, we explore the Edge of Stability phenomenon, indicating its inapplicability even to Lipschitz continuous convex differentiable functions, leaving its relevance to non-convex non-differentiable neural networks inconclusive. Our analysis exposes misguided interpretations of NGDMs in widely referenced papers and texts due to an overreliance on strong smoothness assumptions, emphasizing the necessity for a nuanced understanding of foundational assumptions in the analysis of these systems.