A Theoretical Explanation for Perplexing Behaviors of Backpropagation-based Visualizations

TL;DR

This paper provides a theoretical explanation for why Guided Backpropagation and Deconvolutional Networks produce more interpretable but less class-sensitive visualizations in CNNs, attributing it to image recovery processes caused by specific network components.

Contribution

It introduces a theoretical framework explaining the behaviors of GBP and DeconvNet, linking their visualizations to image recovery rather than network decisions.

Findings

GBP and DeconvNet perform partial image recovery unrelated to class decisions.

Backward ReLU and local CNN connections are key causes of the visualizations.

Experiments support the theoretical analysis.

Abstract

Backpropagation-based visualizations have been proposed to interpret convolutional neural networks (CNNs), however a theory is missing to justify their behaviors: Guided backpropagation (GBP) and deconvolutional network (DeconvNet) generate more human-interpretable but less class-sensitive visualizations than saliency map. Motivated by this, we develop a theoretical explanation revealing that GBP and DeconvNet are essentially doing (partial) image recovery which is unrelated to the network decisions. Specifically, our analysis shows that the backward ReLU introduced by GBP and DeconvNet, and the local connections in CNNs are the two main causes of compelling visualizations. Extensive experiments are provided that support the theoretical analysis.