Responsible model deployment via model-agnostic uncertainty learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Trustworthy AI

Background:

• Predicting machine learning (ML) system failures with production data is vital for trustworthy AI.
• Existing methods lack comprehensive risk assessment and uncertainty decomposition.

Purpose of the Study:

• Introduce Risk Advisor, a post-hoc meta-learner for estimating failure risks and uncertainties of any pre-trained black-box classification model.
• Provide actionable insights into failure causes and guide risk mitigation strategies.

Main Methods:

• Developed Risk Advisor, a meta-learner estimating risk scores and decomposing uncertainty into aleatoric and epistemic components.
• Evaluated performance across diverse black-box models, real-world, and synthetic datasets simulating ML failure scenarios.

Main Results:

• Risk Advisor reliably predicts deployment-time failure risks across various ML failure scenarios.
• The system effectively distinguishes between failures due to data variability, shifts, and model limitations.
• Risk Advisor outperforms existing baseline methods in failure risk prediction.

Conclusions:

• Risk Advisor offers a robust solution for assessing and mitigating ML deployment risks.
• Decomposition of uncertainty provides valuable guidance for targeted risk management.
• The tool enhances the reliability and trustworthiness of AI systems in production environments.