Towards Maximizing the Representation Gap between In-Domain & Out-of-Distribution Examples
This addresses a specific bottleneck in uncertainty estimation for OOD detection, which is important for improving model reliability in safety-critical applications like autonomous systems, but it is incremental as it builds on existing DPN methods.
The paper tackled the problem of indistinguishable representations between in-domain and out-of-distribution (OOD) examples in Dirichlet Prior Networks, which compromises OOD detection performance, by proposing a novel loss function to maximize the representation gap, resulting in consistent improvements in OOD detection.
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.