Bayesian Machine Learning

Deep learning has been remarkably effective in many domains, including amongst others computer vision and reinforcement learning. However, two of the major drawbacks of deep learning is that the models are unable to articulate uncertainty and model selection is typically achieved by trial and error. In this project we explore alternatives to standard deep learning approaches, including Bayesian deep learning and belief propagation in graphical models. Particularly pertinent to robotics applications we hope to show that these methods allow for probabilistic data fusion (e.g. from multiple sensors), and can lead to safer robotics practices through uncertainty awareness. Further, using Bayesian model selection may allow for more efficient model design.