Enhancing Deep Neural Network Reliability with Refinement and Calibration
Researchers have developed a new framework called RefCal to improve the reliability of deep neural networks. This framework jointly optimizes accuracy, calibration, and refinement, addressing the common issue where improving calibration can degrade refinement. RefCal utilizes a novel loss function based on supervised contrastive learning to explicitly promote refinement. In tests on the CIFAR-100-LT dataset, RefCal significantly outperformed existing methods in accuracy, refinement, and expected calibration error. AI
IMPACT Enhances DNN reliability by improving confidence estimates, potentially increasing user trust in AI systems.