Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

EXPLOR: Extrapolatory Pseudo-Label Matching for Out-of-distribution Uncertainty Based Rejection

Created by
  • Haebom

Author

Yunni Qu (Department of Computer Science, University of North Carolina at Chapel Hill), James Wellnitz (Eshelman School of Pharmacy, University of North Carolina at Chapel Hill), Dzung Dinh (Department of Computer Science, University of North Carolina at Chapel Hill), Bhargav Vaduri (Department of Computer Science, University of North Carolina at Chapel Hill), Alexander Tropsha (Eshelman School of Pharmacy, University of North Carolina at Chapel Hill), Junier Oliva (Department of Computer Science, University of North Carolina at Chapel Hill)

Outline

EXPLOR is a novel framework that leverages augmented estimated virtual labeling to improve prediction and uncertainty-based rejection of out-of-distribution (OOD) data. It employs multiple MLP heads (one per base model) with shared embeddings and a novel head-specific matching loss to improve OOD performance on augmented data using different base models as virtual labelers. Unlike previous methods that rely on modality-specific augmentation or assume access to OOD data, EXPLOR introduces estimated virtual labeling for latent space augmentation, enabling robust OOD generalization to any real-valued vector data. Unlike previous modality-independent methods that use neural network backbones, EXPLOR is model-agnostic and works effectively on a wide range of OOD generalization models, from simple tree-based models to complex models. We demonstrate that it outperforms state-of-the-art methods in single-source domain generalization settings across diverse datasets.

Takeaways, Limitations

•
Takeaways:
◦
Robust OOD generalization across diverse data types with estimated virtual labeling via latent space augmentation.
◦
Model-independent framework applicable to various basic models.
◦
Achieving SOTA performance in a single-source domain generalization setup.
•
Limitations:
◦
Experimental results limited to single-source domain generalization settings. Multi-source domain generalization performance is not validated.
◦
Further analysis is needed on the generalization performance and optimization of the proposed matching loss function.
◦
Further research is needed to analyze the performance differences between various base models and to develop strategies for selecting the optimal base model.
👍
Made with Slashpage