Extracting Robust On-Manifold Interactions Encoded by Neural Networks
Keywords: Robustness, Out-of-distribution, Neural Networks
TL;DR: This paper proposes a method to extract robust and stable on-manifold interactions among input variables from DNNs.
Abstract: This paper aims to extract faithful interactions between input variables encoded by a deep neural network (DNN). Recent studies (Ren et al., 2023d; Li & Zhang, 2023b) provided lots of mathematical evidence to support that interactions can be roughly considered as primitive inference patterns encoded by a DNN, given that a small number of interactions can accurately explain the network outputs on any randomly masked samples. However, the instability of interactions to small perturbations on the input still hinders people from taking interactions as rigorous primitives for the network inference. Therefore, in this paper, we propose to extract on-manifold interactions, which are shared by different perturbed inputs. The extracted on-manifold interactions can also explain primitives for adversarial vulnerability. The code will be released after the acceptance of the paper.
Primary Area: general machine learning (i.e., none of the above)
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: 3590
Loading