A Theoretical Explanation for Perplexing Behaviors of Backpropagation-based Visualizations

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