Uncertainty Is Not a Safety Net for Clinical VQA, but Can It Anticipate Model Failure?
A new paper published on arXiv investigates the reliability of uncertainty estimation (UE) methods in clinical visual question-answering (VQA) models. The study found that current UE methods do not consistently indicate when model predictions should be trusted, as their quality degrades with model accuracy. However, the research suggests that UE can still serve as a diagnostic tool, reliably anticipating model fragility when subjected to specific perturbations. AI
IMPACT Current uncertainty estimation methods in clinical VQA models are unreliable for predicting failure, but can diagnose fragility, motivating new evaluation approaches.