Probabilistic deep learning

Deep learning is an extremely popular machine learning (artificial intelligence) approach that has been extremely successful across many applications, such as computer vision, natural language processing, medicine and genomics. Despite this tremendous success, most deep learning methods fail to a) provide estimates of uncertainty in their predictions, b) explain and interpret such predictions and c) have theoretical results for their excellent predictive performance. In this project, we will create new probabilistic deep learning methods that will address these limitations. The contributions will be new probabilistic and interpretable deep neural network models, new algorithms for approximate inference in such models and new theory for explaining their empirical performance.