Class-Similarity Based Label Smoothing for Confidence Calibration

Generating confidence calibrated outputs is of utmost importance for the applications deep neural networks in safety-critical decision-making systems. The output a network probability distribution where scores are estimated confidences input belonging to corresponding classes, and hence they represent complete estimate likelihood relative all classes. In this paper, we propose novel form label smoothing improve calibration. Since different classes intrinsic similarities, more similar should result closer values final output. This motivates development new smooth based on similarities with reference class. We adopt similarity measurements, including those that capture feature-based or semantic similarity. demonstrate through extensive experiments, various datasets architectures, our approach consistently outperforms state-of-the-art calibration techniques uniform smoothing.