Automated Model Inference for Gaussian Processes

Gaussian process models (GPMs) are widely regarded as a prominent tool for learning statistical data models that enable interpolation, regression, and classification. These models are typically instantiated by a Gaussian Process with a zero-mean function and a radial basis covariance function. While these default instantiations yield acceptable analytical quality in terms of model accuracy, GPM inference algorithms automatically search for an application-specific model fitting a particular dataset. State-of-the-art methods for automated inference of GPMs are searching the space of possible models in a rather intricate way and thus result in super-quadratic computation time complexity for model selection and evaluation. Since these properties only enable processing small datasets with low statistical versatility, various methods and algorithms using global as well as local approximations have been proposed for efficient inference of large-scale GPMs. While the latter approximation relies on representing data via local sub-models, global approaches capture data’s inherent characteristics by means of an educated sample. In this paper, we investigate the current state-of-the-art in automated model inference for Gaussian processes and outline strengths and shortcomings of the respective approaches. A performance analysis backs our theoretical findings and provides further empirical evidence. It indicates that approximated inference algorithms, especially locally approximating ones, deliver superior runtime performance, while maintaining the quality level of those using non-approximative Gaussian processes.