On the Two Facets to Conquer Wild Out-of-Distribution Detection.

Area of Science:

• Machine Learning
• Computer Vision
• Artificial Intelligence

Background:

• Out-of-distribution (OOD) detection is crucial for open-world classification, identifying data semantically different from in-distribution (ID) data.
• Outlier Exposure (OE) improves OOD detection by including OOD data during training.
• Wild OOD detection, where OOD data contains ID semantics, compromises model reliability but is understudied.

Purpose of the Study:

• To theoretically analyze the challenges of wild OOD detection from instance and distribution perspectives.
• To introduce general solutions for improving OOD detection reliability in the presence of ID semantics within OOD data.
• To develop a unified framework integrating complementary solutions for robust wild OOD detection.

Main Methods:

• Instance facet: A framework to dynamically estimate true ID/OOD indicators from wild OOD data, mitigating mislabeling.
• Distribution facet: A resampling scheme to remove potential ID sub-distributions from wild OOD data, guided by ID distribution.
• Integration of both methods into a unified framework with theoretical guarantees and practical algorithms.

Main Results:

• The proposed methods effectively mitigate the negative impacts of mislabeled instances and misleading ID sub-distributions in wild OOD data.
• Comprehensive empirical evaluations demonstrate superior performance and reliability compared to existing advanced methods.
• The unified framework leverages complementary strengths for enhanced wild OOD detection efficacy.

Conclusions:

• The developed approaches provide robust solutions for the critical issue of wild OOD detection.
• The theoretical analysis and practical algorithms offer significant advancements in open-world classification reliability.
• This work paves the way for more dependable AI systems in complex, real-world scenarios.