Subnetwork ensembling and data augmentation: Effects on calibration

Deep Learning models based on convolutional neural networks are known to be uncalibrated, that is, they are either overconfident or underconfident in their predictions. Safety-critical applications of neural networks, however, require models to be well-calibrated, and there are various methods in the literature to increase model performance and calibration. Subnetwork ensembling is based on the over-parametrization of modern neural networks by fitting several subnetworks into a single network to take advantage of ensembling them without additional computational costs. Data augmentation methods have also been shown to enhance model performance in terms of accuracy and calibration. However, ensembling and data augmentation seem orthogonal to each other, and the total effect of combining these two methods is not well-known; the literature in fact is inconsistent. Through an extensive set of empirical experiments, we show that combining subnetwork ensemble methods with data augmentation methods does not degrade model calibration.