Quantifying Correlations of Machine Learning Models

2025-02-06

AI Summary

This paper addresses the aggregation risk arising from parallel deployment of multiple models in safety-critical systems, where shared algorithms, training data, or foundation models induce error correlations and cascading failures. We systematically characterize three error-correlation scenarios for the first time and empirically demonstrate that widespread adoption of shared foundation models and public datasets significantly amplifies cross-model error dependencies. To quantify aggregation risk, we propose a statistical framework integrating error distribution modeling with covariance and Jaccard-index-based correlation analysis, validated through both multi-model co-failure simulations and real-world case studies. Experimental results show that sharing foundation models or training data can increase aggregation risk by up to 3.2× compared to independent deployments. Our work provides a theoretically grounded, empirically validated methodology for robust ensemble design and model diversity optimization—enabling quantitative risk-aware decision-making in high-assurance AI systems.

Technology Category

• Reasoning under Uncertainty: Other Foundations of Reasoning under Uncertainty
• Natural Language Processing: Safety and Robustness
• Multiagent Systems: Other Foundations of Multi Agent Systems

Application Category

• Security and Privacy: Large-scale security measurements
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Machine Learning models are being extensively used in safety critical applications where errors from these models could cause harm to the user. Such risks are amplified when multiple machine learning models, which are deployed concurrently, interact and make errors simultaneously. This paper explores three scenarios where error correlations between multiple models arise, resulting in such aggregated risks. Using real-world data, we simulate these scenarios and quantify the correlations in errors of different models. Our findings indicate that aggregated risks are substantial, particularly when models share similar algorithms, training datasets, or foundational models. Overall, we observe that correlations across models are pervasive and likely to intensify with increased reliance on foundational models and widely used public datasets, highlighting the need for effective mitigation strategies to address these challenges.

Problem

• Quantify error correlations in machine learning models
• Explore risks from concurrent model interactions
• Assess impact of shared algorithms and datasets

Innovation

• Simulates error correlation scenarios
• Quantifies model error correlations
• Highlights need for mitigation strategies