Why are Saliency Maps Noisy? Cause of and Solution to Noisy Saliency Maps

Citations Over TimeTop 10% of 2019 papers

Abstract

Saliency Map, the gradient of the score function with respect to the input, is the most basic technique for interpreting deep neural network decisions. However, saliency maps are often visually noisy. Although several hypotheses were proposed to account for this phenomenon, there are few works that provide rigorous analyses of noisy saliency maps. In this paper, we first propose a new hypothesis that noise may occur in saliency maps when irrelevant features pass through ReLU activation functions. Then, we propose Rectified Gradient, a method that alleviates this problem through layer-wise thresholding during backpropagation. Experiments with neural networks trained on CIFAR-10 and ImageNet showed effectiveness of our method and its superiority to other attribution methods.

Related Papers

• → Standard and Advanced Backpropagation Models for Image Processing Application in Traffic Engineering(2002)3 cited
• → Standard and Advanced Backpropagation Models for Image Processing Application in Traffic Engineering(2002)9 cited
• → Global Thresholding and Multiple Pass Parsing(1997)47 cited
• → A localized thresholding method based on boundary detection(2004)