Publications

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

Published in ICML, 2022

This paper proposed three explanation metrics to measure the information discarding during the forward propagation of DNNs. People can fairly compare the representation capacity between different layers and between different DNNs based on our metrics.

Recommended citation: Ma, etc. (2022). "Quantification and Analysis of Layer-wise and Pixel-wise Information Discarding" ICML 2022. http://haozhang37.github.io/files/ICML2022-paper1.pdf

Discovering and Explaining the Representation Bottleneck of DNNs

Published in ICLR, 2022

This paper proved that DNNs usually are more likely to encode both low-order interactions and high-order interactions, but fail to encode middle-order interactions. Such a phenomenon is widely shared by different DNNs.

Recommended citation: Deng, etc. (2022). "Discovering and Explaining the Representation Bottleneck of DNNs" ICLR 2022. http://haozhang37.github.io/files/ICLR2022-paper1.pdf

Interpreting and Boosting Dropout from a Game-Theoretic View

Published in ICLR, 2021

This paper proved that the dropout can effectively reduce the significance of interactions modeled by the DNN both theoretically ad=nd experimentally. Besides, an over-fitted DNN usually encodes more interaction significance. Therefore, dropout can effectively alleviate the over-fitting problem of DNNs.

Recommended citation: Zhang, etc. (2021). "Interpreting and Boosting Dropout from a Game-Theoretic View" ICLR 2021. http://haozhang37.github.io/files/ICLR2021-paper1.pdf

Interpreting Multivariate Shapley Interactions in DNNs

Published in AAAI, 2021

This paper defined and quantified the multivariate interaction. Input variables with strong interactions usually form a coalition and reflect prototype features, which are memorized and used by the DNN for inference.

Recommended citation: Zhang, etc. (2021). "Interpreting Multivariate Shapley Interactions in DNNs" AAAI 2021. http://haozhang37.github.io/files/AAAI2021-paper1.pdf

Interpretable Complex-Valued Neural Networks for Privacy Protection

Published in ICLR, 2020

This paper proposed a method to hide the input information into a random phase of Complex-Valued neural networks. In this way, attackers cannot fetch the input information from intermediate-layer features.

Recommended citation: Xiang, etc. (2020). "Interpretable Complex-Valued Neural Networks for Privacy Protection" ICLR 2020. http://haozhang37.github.io/files/ICLR2020-paper1.pdf

Hao Zhang