<ul><li>Deep neural networks used in safety-critical applications are often poorly calibrated and overconfident.</li><li>Sharpness-aware minimization (SAM) can improve calibration by implicitly maximizing the entropy of the predictive distribution.</li><li>A variant of SAM, CSAM, is proposed to further enhance model calibration.</li><li>Experiments on datasets like ImageNet-1K show SAM's effectiveness in reducing calibration error, with CSAM outperforming SAM and other methods.</li></ul>

Towards Understanding The Calibration Benefits of Sharpness-Aware Minimization

Discover more