Compositional uncertainty in deep Gaussian processes

TL;DR

This paper investigates the limitations of mean-field variational inference in deep Gaussian processes, proposing alternative schemes that better preserve the model's ability to discover compositional structures in data.

Contribution

It identifies the drawbacks of mean-field assumptions in DGPs and explores alternative variational inference methods that maintain layer dependencies.

Findings

Mean-field assumptions cause DGP layers to become nearly deterministic.

Alternative variational schemes can better capture compositional structures.

Proposed methods have potential advantages over traditional mean-field approaches.

Abstract

Gaussian processes (GPs) are nonparametric priors over functions. Fitting a GP implies computing a posterior distribution of functions consistent with the observed data. Similarly, deep Gaussian processes (DGPs) should allow us to compute a posterior distribution of compositions of multiple functions giving rise to the observations. However, exact Bayesian inference is intractable for DGPs, motivating the use of various approximations. We show that the application of simplifying mean-field assumptions across the hierarchy leads to the layers of a DGP collapsing to near-deterministic transformations. We argue that such an inference scheme is suboptimal, not taking advantage of the potential of the model to discover the compositional structure in the data. To address this issue, we examine alternative variational inference schemes allowing for dependencies across different layers and discuss their advantages and limitations.