Multi-level Residual Networks from Dynamical Systems View

TL;DR

This paper interprets deep residual networks through the lens of dynamical systems, providing theoretical insights, analyzing lesioning effects, and proposing a new training acceleration method that reduces training time significantly while maintaining accuracy.

Contribution

It introduces a dynamical systems perspective for ResNets, offers theoretical and experimental analysis, and proposes a novel method to accelerate training.

Findings

Reduced training time by over 40% on image classification benchmarks.

Achieved comparable or superior accuracy with the proposed acceleration method.

Provided theoretical insights into the lesioning properties of ResNets.

Abstract

Deep residual networks (ResNets) and their variants are widely used in many computer vision applications and natural language processing tasks. However, the theoretical principles for designing and training ResNets are still not fully understood. Recently, several points of view have emerged to try to interpret ResNet theoretically, such as unraveled view, unrolled iterative estimation and dynamical systems view. In this paper, we adopt the dynamical systems point of view, and analyze the lesioning properties of ResNet both theoretically and experimentally. Based on these analyses, we additionally propose a novel method for accelerating ResNet training. We apply the proposed method to train ResNets and Wide ResNets for three image classification benchmarks, reducing training time by more than 40% with superior or on-par accuracy.