Improving Interpretability by Information Bottleneck Saliency Guided Localization
Hao Zhou (Jiangsu University), Keyang Cheng (Jiangsu University),* Yu Si (Jiangsu University), Liuyang Yan (Jiangsu University)
The 33rd British Machine Vision Conference
Abstract
The saliency map produced by current deep neural network models fails to accurately focus on important regions of an image due to the influence of input noise. In this paper, we propose a deep learning interpretability method based on information bottleneck, which guides the model training by the probability distribution between the saliency map attributed by the information bottleneck and the gradient-based saliency map. This approach corrects the important regions focused by the model from an information-theoretic perspective. Meanwhile, a saliency suppression mechanism is presented to keep the saliency map of the model away from incorrect classification results and close to correct ones. Experiments show that our method can improve the saliency localization of the model while retaining its accuracy. Compared with other state-of-the-art methods, the Average Drop rate improves by 1.57% and 1.43%, and the Average Increase rate improves by 2.18% and 0.18% in the ResNet-50 model and the VGG-16 model, respectively.Video
Citation
@inproceedings{Zhou_2022_BMVC,
author = {Hao Zhou and Keyang Cheng and Yu Si and Liuyang Yan},
title = {Improving Interpretability by Information Bottleneck Saliency Guided Localization},
booktitle = {33rd British Machine Vision Conference 2022, {BMVC} 2022, London, UK, November 21-24, 2022},
publisher = {{BMVA} Press},
year = {2022},
url = {https://bmvc2022.mpi-inf.mpg.de/0605.pdf}
}
Copyright © 2022
The British Machine Vision Conference is organised by