ASAT: Adaptive OOD Detection Framework with Human Feedback
The ASAT (Adaptive Scoring and Thresholding with Human Feedback) framework, recently introduced, seeks to improve out-of-distribution (OOD) detection in critical safety areas. This innovation, outlined in arXiv paper 2505.02299, tackles the shortcomings of current techniques that rely on static scoring functions and thresholds derived solely from in-distribution (ID) data, often resulting in high false positive rates (FPR). By integrating a human-in-the-loop approach, ASAT dynamically adjusts scoring functions and thresholds in response to real-world OOD data, enhancing its adaptability. The paper, classified as a replace-cross type on arXiv, highlights the dangers posed by OOD inputs and underscores the necessity for effective detection, particularly in fields like autonomous systems and medical diagnostics.
Key facts
- ASAT is a human-in-the-loop framework for OOD detection.
- It safely updates scoring functions and thresholds on the fly.
- Addresses high false positive rates in existing methods.
- Targets evolving OOD inputs for adaptivity.
- Paper available on arXiv with ID 2505.02299.
- Announcement type: replace-cross.
- Focus on safety-critical domains.
- Uses real-world OOD inputs for updates.
Entities
Institutions
- arXiv