EXPLOR: Extrapolatory Pseudo-Label Matching for Out-of-distribution Uncertainty Based Rejection
This addresses the challenge of robust OOD generalization for machine learning models, enabling better handling of unseen data without requiring OOD data access, though it appears incremental as it builds on prior pseudo-labeling and augmentation techniques.
The paper tackles the problem of improving prediction and uncertainty-based rejection on out-of-distribution (OOD) points by introducing EXPLOR, a framework that uses extrapolatory pseudo-labeling on latent-space augmentations, achieving superior performance compared to state-of-the-art methods on diverse datasets in single-source domain generalization settings.
EXPLOR is a novel framework that utilizes support-expanding, extrapolatory pseudo-labeling to improve prediction and uncertainty-based rejection on out-of-distribution (OOD) points. EXPLOR utilizes a diverse set of base models as pseudo-labelers on the expansive augmented data to improve OOD performance through multiple MLP heads (one per base model) with shared embedding trained with a novel per-head matching loss. Unlike prior methods that rely on modality-specific augmentations or assume access to OOD data, EXPLOR introduces extrapolatory pseudo-labeling on latent-space augmentations, enabling robust OOD generalization with any real-valued vector data. In contrast to prior modality-agnostic methods with neural backbones, EXPLOR is model-agnostic, working effectively with methods from simple tree-based models to complex OOD generalization models. We demonstrate that EXPLOR achieves superior performance compared to state-of-the-art methods on diverse datasets in single-source domain generalization settings.