Visualizing and Comparing Convolutional Neural Networks

TL;DR

This paper explores the internal mechanisms of CNNs by visualizing their internal representations and comparing different architectures to understand how depth influences their performance and internal information retention.

Contribution

It introduces methods for visualizing CNN internal representations and provides a comparative analysis of CNNs with varying depths to reveal their internal working mechanisms.

Findings

Deeper CNNs retain more complex visual information.

Visualization of internal layers helps understand CNN decision processes.

Deeper architectures show advantages in internal representation complexity.

Abstract

Convolutional Neural Networks (CNNs) have achieved comparable error rates to well-trained human on ILSVRC2014 image classification task. To achieve better performance, the complexity of CNNs is continually increasing with deeper and bigger architectures. Though CNNs achieved promising external classification behavior, understanding of their internal work mechanism is still limited. In this work, we attempt to understand the internal work mechanism of CNNs by probing the internal representations in two comprehensive aspects, i.e., visualizing patches in the representation spaces constructed by different layers, and visualizing visual information kept in each layer. We further compare CNNs with different depths and show the advantages brought by deeper architecture.