Reliable Classifications with Guaranteed Confidence using the Dempster-Shafer Theory of Evidence
Primary Area: probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
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: Reliable ML, Uncertainty Quantification, Conformal Prediction, Dempster-Shafer Theory
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
TL;DR: We develop a method for reliable classifications in presence of epistemic uncertainty using the Dempster-Shafer theory of evidence in combination with conformal prediction.
Abstract: Reliably capturing predictive uncertainty is indispensable for the deployment of machine learning (ML) models in safety-critical domains. The most commonly used approaches to uncertainty quantification are, however, either computationally costly in inference or incapable of capturing different types of uncertainty (i.e., aleatoric and epistemic). In this paper, we tackle this issue using the Dempster-Shafer theory of evidence, which only recently gained attention as a tool to estimate uncertainty in ML. By training a neural network to return a generalized probability measure and combining it with conformal prediction, we obtain set predictions with guaranteed user-specified confidence. We test our method on various datasets and empirically show that it reflects uncertainty more reliably than a calibrated classifier with softmax output, since our approach yields smaller and hence more informative prediction sets at the same bounded error level in particular for samples with high epistemic uncertainty. In order to deal with the exponential scaling inherent to classifiers within Dempster-Shafer theory, we introduce a second approach with reduced complexity, which also returns smaller sets than the comparative method, even on large classification tasks with more than 40 distinct labels. Our results indicate that the proposed methods are promising approaches to obtain reliable and informative predictions in the presence of both aleatoric and epistemic uncertainty in only one forward-pass through the network.
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: 4945
Loading