Learning Interpretable and Influential Directions with Signal Vectors and Uncertainty Region Alignment

ICLR 2025 Conference Submission4691 Authors

25 Sept 2024 (modified: 28 Nov 2024)ICLR 2025 Conference SubmissionEveryoneRevisionsBibTeXCC BY 4.0
Keywords: latent space, interpretability, concepts, directions, signals, patterns, distractors
TL;DR: The proposed unsupervised method identifies a pair of latent space directions (filter and signal) with the first being able to answer questions of interpretability and the second to answer questions of concept influence on model's predictions
Abstract: Latent space directions have played a key role in understanding, debugging, and fixing deep learning models. Concepts are often encoded in distinct feature space directions, and evaluating impact of these directions on the model's predictions, highlights their importance in the decision-making process. Additionally, recent studies have shown that penalizing directions associated with spurious artifacts during training can force models to unlearn features irrelevant to their prediction task. Identifying these directions, therefore, provides numerous benefits, including a deeper understanding of the model's strategy, fostering trust, and enabling model correction and improvement. We introduce a novel unsupervised approach utilizing signal vectors and uncertainty region alignment to discover latent space directions that meet two key debugging criteria: significant influence on model predictions and high level of interpretability. To our knowledge, this method is the first of its kind to uncover such directions, leveraging the inherent structure of the feature space and the knowledge encoded in the deep network. We validate our approach using both synthetic and real-world benchmarks, demonstrating that the discovered directions effectively fulfill the critical debugging criteria.
Primary Area: interpretability and explainable AI
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/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
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: 4691
Loading