Agentic data science with the Data Agent Kit
cloud.google.com·6h ago
AI Research
arXiv CS.AI·Jul 27·5 min read
Self-Poisoning in Adaptive Out-of-Distribution Detection: A Sharp-Threshold Theory and Certified Label-Free Calibration
TL;DR
A new method for adaptive out-of-distribution (OOD) detection reveals a critical threshold for memory bank contamination, leading to detector failure. The approach includes a certified admission gate that mitigates contamination effects while maintaining low false positive rates.
✦ Why It Matters
Implement certified admission gates in your OOD detection systems to enhance robustness against contamination.
Key Takeaways
- Utilize a certified admission gate to prevent memory bank contamination.
- Monitor the impurity level of your memory bank to avoid detector collapse.
- Apply CDC to maintain nominal false positive rates under data drift.
- Understand the limitations of label-free detection methods in distinguishing drift from contamination.
How It Works
The study employs a generalized Pólya urn model to analyze memory bank impurity, establishing a critical threshold for contamination. When impurity exceeds this threshold, the detector's performance collapses.
The certified admission gate acts as a safeguard, using a static reserve to prevent feedback loops that could lead to contamination.