Deep Gaussian Processes for Large Datasets

Big data and Bayesian non-parametric modelling constitute two of the most important foci of modern machine learning research. In this preliminary work we propose a neat solution for combining the aforementioned domains into a single principled framework based on Gaussian processes. Speficically, we invistigate algorithms for training deep generative models with hidden layers connected with non-linear Gaussian process (GP) mappings. Building on recent developments on (stochastic) variational approximations, the models are fitted on massive data and the hidden variables are marginalised out in a Bayesian manner to allow for efficient propagation of the uncertainty throughout the network of variables.