Is Each Layer Non-trivial in CNN?

TL;DR

This paper investigates whether all layers in CNNs, specifically ResNet, are essential by replacing kernels with zeros and observing minimal performance loss, revealing some layers are trivial.

Contribution

It introduces a method to evaluate layer importance in CNNs by kernel ablation, showing many layers in ResNet are trivial and not critical for performance.

Findings

Some convolution kernels are trivial and can be replaced with zeros without performance loss

Not all layers in ResNet are essential, challenging assumptions about layer importance

The approach provides a new way to analyze and potentially prune CNNs

Abstract

Convolutional neural network (CNN) models have achieved great success in many fields. With the advent of ResNet, networks used in practice are getting deeper and wider. However, is each layer non-trivial in networks? To answer this question, we trained a network on the training set, then we replace the network convolution kernels with zeros and test the result models on the test set. We compared experimental results with baseline and showed that we can reach similar or even the same performances. Although convolution kernels are the cores of networks, we demonstrate that some of them are trivial and regular in ResNet.