Sparse Gaussian processes using power expectation propagation for GPyTorch

This repository implements the Power EP algorithm for Sparse GPs as described by Bui & Yan & Turner (2017). The code tries to follow the original implementation by Bui and has comments referencing the original code for comparison’s sake.

Requirements

This was implemented against torch==1.10.2, gpytorch==1.7.0.

Limitations

• The code does not implement hyperparameter training through gradients of the approximation of the evidence;
• No support for non-zero prior mean;
• No support for batched GPs;
• No support for multi-output GPs.