Predictive Uncertainty Quantification of Deep Neural Networks using Dirichlet Distributions

Advancements in deep neural networks have made it a prominent approach for most of the complex computer vision tasks. A key aspect for the deployment of deep neural networks in several applications, like automotive and medical, has been its ability to estimate its uncertainty. A recent leading approach is using Dirichlet distributions to model the uncertainty, which results in real-time estimation of uncertainty. The intermediate layer variational inference has also been a promising approach to-enable real-time estimation of uncertainty, beating state-of-the-art approaches. In this work we introduce the incorporation of both approaches in order to improve the reliability of uncertainty estimation whilst maintaining the real-time capability. Our experiments on the Cityscapes dataset for the task of semantic segmentation showed a significant boost in the deep neural network's uncertainty estimation capability, whilst also improving its segmentation performance.