Obtaining reliable and accurate quantification of uncertainty estimates from\ndeep neural networks is important in safety-critical applications. A\nwell-calibrated model should be accurate when it is certain about its\nprediction and indicate high uncertainty when it is likely to be inaccurate.\nUncertainty calibration is a challenging problem as there is no ground truth\navailable for uncertainty estimates. We propose an optimization method that\nleverages the relationship between accuracy and uncertainty as an anchor for\nuncertainty calibration. We introduce a differentiable accuracy versus\nuncertainty calibration (AvUC) loss function that allows a model to learn to\nprovide well-calibrated uncertainties, in addition to improved accuracy. We\nalso demonstrate the same methodology can be extended to post-hoc uncertainty\ncalibration on pretrained models. We illustrate our approach with mean-field\nstochastic variational inference and compare with state-of-the-art methods.\nExtensive experiments demonstrate our approach yields better model calibration\nthan existing methods on large-scale image classification tasks under\ndistributional shift.\n