Quantification and Analysis of Layer-wise and Pixel-wise Information Discarding

Haotian Ma, Hao Zhang, Fan Zhou, Yinqing Zhang, Quanshi Zhang
Proceedings of the 39th International Conference on Machine Learning, PMLR 162:14664-14698, 2022.

Abstract

This paper presents a method to explain how the information of each input variable is gradually discarded during the forward propagation in a deep neural network (DNN), which provides new perspectives to explain DNNs. We define two types of entropy-based metrics, i.e. (1) the discarding of pixel-wise information used in the forward propagation, and (2) the uncertainty of the input reconstruction, to measure input information contained by a specific layer from two perspectives. Unlike previous attribution metrics, the proposed metrics ensure the fairness of comparisons between different layers of different DNNs. We can use these metrics to analyze the efficiency of information processing in DNNs, which exhibits strong connections to the performance of DNNs. We analyze information discarding in a pixel-wise manner, which is different from the information bottleneck theory measuring feature information w.r.t. the sample distribution. Experiments have shown the effectiveness of our metrics in analyzing classic DNNs and explaining existing deep-learning techniques. The code is available at https://github.com/haotianSustc/deepinfo.

Cite this Paper

BibTeX
@InProceedings{pmlr-v162-ma22b, title = {Quantification and Analysis of Layer-wise and Pixel-wise Information Discarding}, author = {Ma, Haotian and Zhang, Hao and Zhou, Fan and Zhang, Yinqing and Zhang, Quanshi}, booktitle = {Proceedings of the 39th International Conference on Machine Learning}, pages = {14664--14698}, year = {2022}, editor = {Chaudhuri, Kamalika and Jegelka, Stefanie and Song, Le and Szepesvari, Csaba and Niu, Gang and Sabato, Sivan}, volume = {162}, series = {Proceedings of Machine Learning Research}, month = {17--23 Jul}, publisher = {PMLR}, pdf = {https://proceedings.mlr.press/v162/ma22b/ma22b.pdf}, url = {https://proceedings.mlr.press/v162/ma22b.html}, abstract = {This paper presents a method to explain how the information of each input variable is gradually discarded during the forward propagation in a deep neural network (DNN), which provides new perspectives to explain DNNs. We define two types of entropy-based metrics, i.e. (1) the discarding of pixel-wise information used in the forward propagation, and (2) the uncertainty of the input reconstruction, to measure input information contained by a specific layer from two perspectives. Unlike previous attribution metrics, the proposed metrics ensure the fairness of comparisons between different layers of different DNNs. We can use these metrics to analyze the efficiency of information processing in DNNs, which exhibits strong connections to the performance of DNNs. We analyze information discarding in a pixel-wise manner, which is different from the information bottleneck theory measuring feature information w.r.t. the sample distribution. Experiments have shown the effectiveness of our metrics in analyzing classic DNNs and explaining existing deep-learning techniques. The code is available at https://github.com/haotianSustc/deepinfo.} }
Endnote
%0 Conference Paper %T Quantification and Analysis of Layer-wise and Pixel-wise Information Discarding %A Haotian Ma %A Hao Zhang %A Fan Zhou %A Yinqing Zhang %A Quanshi Zhang %B Proceedings of the 39th International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2022 %E Kamalika Chaudhuri %E Stefanie Jegelka %E Le Song %E Csaba Szepesvari %E Gang Niu %E Sivan Sabato %F pmlr-v162-ma22b %I PMLR %P 14664--14698 %U https://proceedings.mlr.press/v162/ma22b.html %V 162 %X This paper presents a method to explain how the information of each input variable is gradually discarded during the forward propagation in a deep neural network (DNN), which provides new perspectives to explain DNNs. We define two types of entropy-based metrics, i.e. (1) the discarding of pixel-wise information used in the forward propagation, and (2) the uncertainty of the input reconstruction, to measure input information contained by a specific layer from two perspectives. Unlike previous attribution metrics, the proposed metrics ensure the fairness of comparisons between different layers of different DNNs. We can use these metrics to analyze the efficiency of information processing in DNNs, which exhibits strong connections to the performance of DNNs. We analyze information discarding in a pixel-wise manner, which is different from the information bottleneck theory measuring feature information w.r.t. the sample distribution. Experiments have shown the effectiveness of our metrics in analyzing classic DNNs and explaining existing deep-learning techniques. The code is available at https://github.com/haotianSustc/deepinfo.
APA
Ma, H., Zhang, H., Zhou, F., Zhang, Y. & Zhang, Q.. (2022). Quantification and Analysis of Layer-wise and Pixel-wise Information Discarding. Proceedings of the 39th International Conference on Machine Learning, in Proceedings of Machine Learning Research 162:14664-14698 Available from https://proceedings.mlr.press/v162/ma22b.html.

Related Material