The Uncertainty-Perception Tradeoff
Primary Area: generative models
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: uncertainty-perception tradeoff, uncertainty quantification, statistical estimation theory, image restoration
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
TL;DR: We prove mathematically there is a tradeoff between perceptual quality and uncertainty in generative models for restoration tasks.
Abstract: Generative models have achieved groundbreaking performance in restoration tasks and inverse problems, producing results that are often indistinguishable from real data. Yet these models are also known to produce hallucinations, or artifacts that are not present in the original input, raising concerns about the uncertainty of the models' predictions. In this paper we study this phenomenon, employing information-theory tools to reveal a fundamental tradeoff between perception and uncertainty. Our mathematical analysis shows that as perceptual quality increases, so does the uncertainty of a restoration algorithm as quantified by error entropy. We derive and illustrate the behavior of the uncertainty-perception function, showcasing both local and global bounds that define the the feasible region of the tradeoff. Furthermore, we revisit a well-known relation between estimation distortion and uncertainty and generalize its scope to include perception quality, thereby shedding new light on the well-established perception-distortion tradeoff. Our work offers a principled analysis of uncertainty, highlighting its interplay with perception and the limitations of generative models in restoration 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: 1821
Loading