Deep neural networks have shown great success in prediction quality while\nreliable and robust uncertainty estimation remains a challenge. Predictive\nuncertainty supplements model predictions and enables improved functionality of\ndownstream tasks including embedded and mobile applications, such as virtual\nreality, augmented reality, sensor fusion, and perception. These applications\noften require a compromise in complexity to obtain uncertainty estimates due to\nvery limited memory and compute resources. We tackle this problem by building\nupon Monte Carlo Dropout (MCDO) models using the Axolotl framework;\nspecifically, we diversify sampled subnetworks, leverage dropout patterns, and\nuse a branching technique to improve predictive performance while maintaining\nfast computations. We conduct experiments on (1) a multi-class classification\ntask using the CIFAR10 dataset, and (2) a more complex human body segmentation\ntask. Our results show the effectiveness of our approach by reaching close to\nDeep Ensemble prediction quality and uncertainty estimation, while still\nachieving faster inference on resource-limited mobile platforms.\n