Adaptive Confidence Regularization for Multimodal Failure Detection

Researchers propose Adaptive Confidence Regularization (ACR), a new framework for detecting failures in multimodal AI systems used in critical applications like autonomous vehicles and medical diagnostics. The approach uses confidence degradation detection and synthetic failure generation to improve reliability of AI predictions in high-stakes scenarios.

• →ACR addresses the previously unexplored problem of failure detection in multimodal AI systems.
• →The framework identifies confidence degradation where multimodal predictions show lower confidence than individual components.
• →Multimodal Feature Swapping generates synthetic failure cases to improve training robustness.
• →Testing across four datasets and three modalities demonstrates consistent performance improvements.
• →The research targets critical applications including autonomous vehicles and medical diagnostics.