Towards Maximizing the Representation Gap between In-Domain & Out-of-Distribution Examples

TL;DR

This paper introduces a new loss function for Dirichlet Prior Networks to enhance the distinction between in-domain and out-of-distribution examples, significantly improving OOD detection accuracy.

Contribution

It proposes a novel loss function that maximizes the representation gap between in-domain and OOD examples in DPN models, addressing a key limitation.

Findings

Improved OOD detection performance across multiple experiments

Enhanced separation of in-domain and OOD representations

Consistent results demonstrating the effectiveness of the proposed loss

Abstract

Among existing uncertainty estimation approaches, Dirichlet Prior Network (DPN) distinctly models different predictive uncertainty types. However, for in-domain examples with high data uncertainties among multiple classes, even a DPN model often produces indistinguishable representations from the out-of-distribution (OOD) examples, compromising their OOD detection performance. We address this shortcoming by proposing a novel loss function for DPN to maximize the \textit{representation gap} between in-domain and OOD examples. Experimental results demonstrate that our proposed approach consistently improves OOD detection performance.