A Derandomization Framework for Structure Discovery: Applications in Neural Networks and Beyond

TL;DR

This paper introduces a derandomization framework for discovering neural network structures, applicable to various architectures and training methods, with implications for optimization and data embedding techniques.

Contribution

It presents a derandomization lemma that explains structure discovery in neural networks under weak assumptions, extending previous analyses to more general settings.

Findings

The derandomization lemma shows convergence to zero weights under mild conditions.

Application of the framework to MAXCUT approximation and Johnson-Lindenstrauss embeddings.

Supports structure discovery in neural networks trained with any SOSP-attaining method.

Abstract

Understanding the dynamics of feature learning in neural networks (NNs) remains a significant challenge. The work of (Mousavi-Hosseini et al., 2023) analyzes a multiple index teacher-student setting and shows that a two-layer student attains a low-rank structure in its first-layer weights when trained with stochastic gradient descent (SGD) and a strong regularizer. This structural property is known to reduce sample complexity of generalization. Indeed, in a second step, the same authors establish algorithm-specific learning guarantees under additional assumptions. In this paper, we focus exclusively on the structure discovery aspect and study it under weaker assumptions, more specifically: we allow (a) NNs of arbitrary size and depth, (b) with all parameters trainable, (c) under any smooth loss function, (d) tiny regularization, and (e) trained by any method that attains a second-order stationary point (SOSP), e.g.\ perturbed gradient descent (PGD). At the core of our approach is a key $\textit{derandomization}$ lemma, which states that optimizing the function $\mathbb{E}_{\mathbf{x}} \left[g_{\theta}(\mathbf{W}\mathbf{x} + \mathbf{b})\right]$ converges to a point where $\mathbf{W} = \mathbf{0}$, under mild conditions. The fundamental nature of this lemma directly explains structure discovery and has immediate applications in other domains including an end-to-end approximation for MAXCUT, and computing Johnson-Lindenstrauss embeddings.