everyone
since 04 Oct 2024">EveryoneRevisionsBibTeXCC BY 4.0
Reliable prediction is an essential requirement for deep neural models that are deployed in open environments, where both covariate and semantic out-of-distribution (OOD) data arise naturally. Recent studies have formulated and pursued two problems named OOD generalization and detection independently, where the former aims to correctly recognize covariate shifts while the latter focuses on rejecting semantic shifts. However, existing methods are misaligned with real-world applications in two aspects. First, in practice, to make safe decisions, a reliable model should accept correctly recognized inputs while rejecting both those misclassified covariate-shifted and semantic-shifted examples. Second, considering the potential existing trade-off between rejecting different failure cases, more convenient, controllable, and flexible unknown rejection approaches are needed. To meet the above requirements, we propose a novel and elegantly simple unknown rejection framework to unify and facilitate classification with rejection under both covariate and semantic shifts. Our key insight is that by separating and consolidating failure-specific reliability knowledge with low-rank adapters and then integrating them, we can enhance the unknown rejection ability effectively and flexibly. Extensive experiments demonstrate the superiority of our framework.