Reduced Robust Random Cut Forest for Out-of-Distribution Detection in Machine Learning Models

Overview

When ML-based regressors receive inputs with significantly different statistical properties from their training data, predictions can be erroneous or hazardous. This work introduces a novel Reduced Robust Random Cut Forest (RRRCF) data structure for detecting out-of-distribution inputs — critical for safe deployment of autonomous systems.

Key Contributions

• Novel RRRCF algorithm that works on both small and large datasets
• Runtime OOD detection for ML regressors — not just classifiers
• Applicable as a runtime assurance monitor for learning-enabled controllers in autonomous vehicles

Relevance to Assured Autonomy

This directly addresses the AA program’s need for runtime monitoring of learning-enabled components. When a deployed autonomous system encounters conditions outside its training distribution, the RRRCF detector flags the anomaly, triggering fallback safety behaviors.