New taxonomy details neural network failure modes in closed-loop control systems

This paper introduces a taxonomy for failure modes when neural networks are deployed in closed-loop control systems, where the network's actions alter the data distribution it was trained on. It details issues like Closed-Loop Identification Error, Action Error, and Tracking Error, and explains how offline training can lead to performance degradation due to local loss geometry. The research proposes design policies for time-scale separated online adaptation to mitigate these distribution shifts, validated on a nonlinear Duffing oscillator, offering practical guidance for deploying ML in safety-critical feedback systems. AI

IMPACT Provides a framework for understanding and mitigating deployment failures of ML models in safety-critical feedback systems.

RANK_REASON Academic paper detailing a new taxonomy and mitigation strategies for machine learning in control systems. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• Machine Learning
• Neural Networks

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →