Similarity-as-Evidence: Calibrating Overconfident VLMs for Interpretable and Label-Efficient Medical Active Learning
The Similarity-as-Evidence (SaE) framework recalibrates overconfident Vision-Language Models for medical active learning by treating similarity scores as evidence rather than deterministic predictions. It introduces a Similarity Evidence Head to quantify model certainty, decomposes uncertainty into knowledge gaps (vacuity) and decision conflicts (dissonance), and uses a dual-factor acquisition strategy to select the most informative samples, leading to improved accuracy and calibration.