Tackling the Noisy Elephant in the Room: Label Noise-robust Out-of-Distribution Detection via Loss Correction and Low-rank Decomposition

ICLR 2026 Conference Submission21298 Authors

Keywords: out-of-distribution detection, noisy labels, robust learning

Abstract:

Robust out-of-distribution (OOD) detection is an indispensable component of modern artificial intelligence (AI) systems, especially in safety-critical applications where models must identify inputs from unfamiliar classes not seen during training. While OOD detection has been extensively studied in the machine learning literature—with both post hoc and training-based approaches—its effectiveness under noisy training labels remains underexplored. Recent studies suggest that label noise can significantly degrade OOD performance, yet principled solutions to this issue are lacking.
In this work, we demonstrate that directly combining existing label noise-robust methods with OOD detection strategies is insufficient to address this critical challenge. To overcome this, we propose a $\textit{robust}$ OOD detection framework designed to $\textit{cleanse}$ feature embeddings, thereby mitigating the adverse effects of noisy labels on OOD performance. Towards this, we introduce an end-to-end training strategy that integrates loss correction methods from the noisy-label learning literature with low-rank and sparse decomposition techniques from signal processing. Building on this strategy, we derive a novel metric that quantifies the “OOD-ness” content within training data, which in turn leads to a label noise-robust OOD detection scoring technique.
Extensive experiments on both synthetic and real-world datasets demonstrate that our method significantly outperforms the state-of-the-art OOD detection techniques, particularly under severe noisy label settings.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 21298