Power of Augmented Replicas in Out-Of-Distribution Detection

IMANOL ALLENDE, Nicholas Mc Guire, Carles Hernandez

Keywords: Data Augmentation, Out-of-Distribution Detection (OOD), trustworthy ML

TL;DR: We propose using data augmentation during the inference stage to improve out-of-distribution detection by transforming different copies of the same image and evaluating the consensus of the detectors.

Abstract:

Data augmentation is widely used in machine learning to enhance training datasets by introducing minor variations to the original data, traditionally aiming to prevent overfitting and improve model performance. This paper explores a novel application of data augmentation during the inference stage to enhance out-of-distribution (OOD) detection. The proposed method involves replicating the inference image multiple times, applying various transformation techniques to each replica, and then evaluating the detectors using these augmented images. The effectiveness of this approach is assessed across different detectors, models, and datasets, demonstrating its potential to improve OOD detection capabilities.

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.

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: 9582