Multi-level Residual Networks from Dynamical Systems View
This work addresses the need for faster training of ResNets in computer vision and NLP, offering a practical improvement for researchers and practitioners, though it is incremental as it builds on existing dynamical systems interpretations.
The paper tackled the problem of understanding and accelerating deep residual networks (ResNets) by analyzing their lesioning properties from a dynamical systems view, resulting in a novel training method that reduces training time by over 40% while maintaining or improving accuracy on image classification benchmarks.
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.