Towards calibrated and scalable uncertainty representations for neural networks