Peacock: Multi-Objective Optimization for Deep Neural Network Calibration

Dexter Neo, Tsuhan Chen

Keywords: Deep Neural Network Calibration, Uncertainty Calibration, Robustness, Safety, Out-of-Distribution

TL;DR: We propose Peacock a novel multi-objective calibration framework motivated by the strengths of recent contributions. We demonstrate how to effectively combine different methods, into a fast and well-optimized calibration algorithm.

Abstract:

The rapid adoption of deep neural networks underscores an urgent need for models to be safe, trustworthy and well-calibrated. Despite recent advancements in network calibration, the optimal combination of techniques remains relatively unexplored. By framing the task as a multi-objective optimization problem, we demonstrate that combining state-of-the-art methods can further boost calibration performance. We feature a total of seven state-of-the-art calibration algorithms and provide both theoretical and empirical motivation for their equal and weighted importance unification. We conduct experiments on both in and out-of-distribution computer vision and natural language benchmarks, investigating the speeds and contributions of different components.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

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.

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: 1888