Classification with Conceptual Safeguards

Published: 16 Jan 2024, Last Modified: 11 Feb 2024ICLR 2024 posterEveryoneRevisionsBibTeX
Primary Area: societal considerations including fairness, safety, privacy
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: classification, interpretability, explainability, concept bottleneck models, uncertainty, selective classification
Submission Guidelines: I certify that this submission complies with the submission instructions as described on
TL;DR: We introduce conceptual safeguards, a method for selective classification in concept bottleneck models that can reliably promote safety and interpretability in automation tasks
Abstract: Machine learning models are often used to automate routine tasks. In settings where mistakes are costly, we can trade off accuracy for coverage by abstaining from making a prediction on instances for which the model is uncertain. In this work, we present a new approach to selective classification in deep learning with concepts. Our approach constructs a concept bottleneck model where the front-end model can make predictions given soft concepts and leverage concept confirmation to improve coverage and performance under abstention. We develop techniques to propagate uncertainty and identify concepts for confirmation. We evaluate our approach on real-world and synthetic datasets, showing that it can improve coverage while maintaining performance across a range of tasks.
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: 2858