W-DOE: Wasserstein Distribution-Agnostic Outlier Exposure

Area of Science:

• Machine Learning
• Computer Vision
• Artificial Intelligence

Background:

• Open-world classification models require robust out-of-distribution (OOD) detection capabilities.
• Outlier Exposure (OE) enhances models using auxiliary OOD data but may suffer from unrepresentative OOD data.
• Existing OE methods can introduce bias, limiting practical OOD detection performance.

Purpose of the Study:

• To propose a novel OE-based learning method, Wasserstein Distribution-agnostic Outlier Exposure (W-DOE), for improved OOD detection.
• To enhance model robustness by expanding the coverage of training-time OOD data.
• To provide theoretical guarantees for open-world settings and improved generalization for unseen OOD cases.

Main Methods:

• Introduced Wasserstein Distribution-agnostic Outlier Exposure (W-DOE), a theoretically sound and experimentally superior OE method.
• Developed Implicit Data Synthesis (IDS), a novel approach to generate additional OOD data by perturbing model parameters.
• Implemented a general learning framework to optimize synthesized OOD data for maximum model benefit, measured by the Wasserstein metric.

Main Results:

• W-DOE demonstrated superior performance against state-of-the-art methods across various OOD detection benchmarks.
• The proposed IDS approach effectively expands OOD data coverage, leading to fewer unseen OOD cases during deployment.
• Broader OOD coverage, as ensured by W-DOE, resulted in reduced estimation errors and improved generalization for real-world OOD scenarios.

Conclusions:

• W-DOE offers a theoretically grounded and empirically validated approach to enhance OOD detection in open-world environments.
• The method's ability to synthesize diverse OOD data via IDS significantly improves model adaptability to novel OOD data.
• W-DOE provides provable guarantees, ensuring more reliable and accurate OOD detection performance in practical applications.