MAXENT LOSS: CONSTRAINED MAXIMUM ENTROPY FOR CALIBRATING DEEP NEURAL NETWORKS

Dexter Neo, Stefan Winkler

Keywords: Calibration, Out-of-distribution, Loss function, Machine learning safety, Overconfidence, Robustness, Distribution shifts

TL;DR: A novel loss function involving constraints, used to improve model calibration on OOD data.

Abstract: Miscalibration distorts our interpretation between a model's confidence and correctness, making them unreliable for real-world deployment. In general, we want dependable and meaningful probabilistic estimations of our model's uncertainty, which are essential in real-world applications. This may include inputs that are from out-of-distributions (OOD), which can be widely different from the given training distribution. Motivated by the Principle of Maximum Entropy, we show that -- compared to conventional cross-entropy loss and focal loss -- training neural networks with additional statistical constraints can help improve neural calibration whilst retaining recognition accuracy. We evaluate our method extensively on different augmented and in-the-wild OOD computer vision datasets and show that our Maxent loss achieves state-of-the-art calibration in all cases. Our code will be made available upon acceptance.

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.

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: Yes

Please Choose The Closest Area That Your Submission Falls Into: Deep Learning and representational learning