Scalable and automated inference for gaussian process models

Gaussian processes (GPs) are widely used in the Bayesian approach to supervised learning. Their ability to provide rich priors over functions is highly desirable for modeling real-world problems. Unfortunately, there exist two big challenges when doing Bayesian inference (i.e., learning the posteriors over functions) for GP models. The first is analytical intractability: The posteriors cannot be computed in closed- form when non-Gaussian likelihoods are employed. The second is scalability: The inference procedures often cannot be applied to large datasets due to their prohibitive computational costs. In this thesis, I develop practical variational inference methods to address the first challenge. Moreover, I introduce three GP models to deal with the second challenge. First, I focus on the analytical intractability challenge starting with the Gaussian process regression networks (GPRN), an expressive multi-output model with adaptive, input-dependent correlations. I derive a variational inference method with two different variational distributions to approximate the true posterior of GPRN. While one distribution is a standard Gaussian, the other is a Gaussian mixture which can capture more complex, multimodal posteriors. Both distributions are shown to be statistically efficient, requiring only a linear number of parameters to represent their inherent covariance matrices. Experimental results demonstrate clear benefits of having a multimodal variational approximation in GPRN. Next, I use the same two variational distributions to address the analytical in- tractability challenge for a large class of GP models. I show that the aforementioned statistical efficiency also stands for members of this class. I further prove that the gradients required for variational learning can either be approximated efficiently or computed analytically, regardless of the likelihood functions of the models. Based on these insights, I develop an automated variational inference method for GP models with general likelihoods. The method allows easy investigation of existing or new models without having to derive model-specific inference algorithms. I then turn to the scalability challenge, focusing on single-output and multi- output regression. The underpinning technique here is sparse GP - a GP augmented with so-called inducing points/variables that lead to lower computational demands. For single-output regression, I introduce a mixture-of-experts model (FGP) where the experts are independent sparse GPs each having their own inducing variables. Their inducing inputs further define a partitioning structure of the input space, allowing an efficient inference scheme in which computation is carried out locally by the experts. FGP can thus be K2 time faster and use K2 less memory than previous GP models, where K is the number of experts. For multi-output regression, I introduce the collaborative multi-output Gaussian process model (COGP) where the outputs are linear combinations of independent sparse GPs. Their inducing points are represented as global variables which correlate the outputs for joint learning. The variables are then exploited to derive a stochastic variational inference method that can deal with a much larger number of inputs and outputs compared to previous models. Superior empirical performance of FGP and COGP is demonstrated through extensive experiments on various real-world datasets