A2047
Title: Certifiable perception: Generating safety maps via conformal segmentation
Authors: Rui Luo - City University of Hong Kong (Hong Kong) [presenting]
Abstract: In safety-critical domains such as medical imaging and autonomous driving, standard image segmentation models often fail to provide reliable uncertainty quantification, leading to dangerous false negatives. While existing Conformal Risk control methods offer marginal coverage guarantees, they often perform inconsistently across diverse images, providing excessive coverage for simple inputs while failing to protect complex ones. Conformal Risk Adaptation (CRA) is a novel framework designed to enhance Conditional Risk control in segmentation by leveraging Adaptive score functions that adjust prediction sets based on the difficulty of individual images. Furthermore, Conditional Optimal Adaptive Thresholding (COAT) is presented as an end-to-end differentiable optimization approach that learns to predict the Optimal, image-specific threshold, effectively minimizing the Risk gap. Experiments demonstrate that these methods provide valid marginal Risk control while significantly improving consistency in Conditional Risk performance, offering a robust, principled solution for high-stakes perception tasks.
