Causal-based Analysis on Credibility of Feedforward Neural Network

Pengzhao Quan, Zhengxi Li, Hongsheng Ma, Yunlin Lei, Hanzhu Jia, Lanyu Gao, Xu Yang

Primary Area: causal reasoning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: causal effect, causal relation, feedforward neural network

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Abstract: Feedforward neural network (FNN) has been widely used in various fields. However, its credibility in some risk-sensitive fields remains questionable. The complex causal relations between neurons in the input layer is hard to be observed. These causal relations affect the credibility of FNN directly or indirectly. We transform FNN into a causal structure with different causal relations in the input layer. In order to analyze the causal structure accurately, we categorize it into three causal sub-structures based on different causal relations between input and output neurons. With different levels of intervention, we analyze the causal effect by calculating average treatment effect for each neuron in the input layer. We conduct experiments in the field of pediatric ophthalmology. The results demonstrate the validity of our causal-based analysis method.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 7579