Iterative Search Attribution for Deep Neural Networks

Zhiyu Zhu, Huaming Chen, Jiayu Zhang, Xinyi Wang, Zhibo Jin, Jason Xue, Jun Shen

Keywords: Attribution, XAI, interpretability

Abstract: Deep neural networks (DNNs) have achieved state-of-the-art performance in a number of application areas. However, to ensure the reliability of a DNN model and achieve a desired level of trustworthiness, it is critical to enhance the interpretability in terms of the model inputs and outputs. Attribution methods are an effective means of Explainable Artificial Intelligence (XAI) research. However, the interpretability of existing attribution algorithms varys depending on the choice of reference point, the quality of constructed adversarial samples, or the applicability of gradient constraints in specific tasks. To effectively and thoroughly explore the attribution integration paths, in this paper, inspired by the iterative generation of high-quality samples in the diffusion model, we propose an Iterative Search Attribution (ISA) method to achieve more accurate attribution by distinguishing the importance of samples during gradient ascent and descent and clipping the relatively unimportant features in the model. Specifically, we introduce a scale parameter during the iterative process to ensure that the parameters in the next iteration are always more significant than the parameters in the current iteration. Comprehensive experimental results show that our method has superior results in image recognition interpretability tasks compared with other sota baselines. Our code is available at: https://anonymous.4open.science/r/ISA-6F6B

Primary Area: visualization or interpretation of learned representations

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.

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

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: 5128