zphilip / streaming_sparse_gp Goto Github PK

This repository contains an implementation of several online/streaming sparse GP approximations for regression and classification (Bui, Nguyen and Turner, NIPS 2017). In particular, osvgp.py implements the uncollapsed variational free-energy for regression and classification, and osgpr.py implements the collapsed variational free-energy and Power-EP energy for the regression case.

We also provide an implementation of the collapsed batch Power-EP sparse approximation of Bui, Yan and Turner (2017).

The code was tested using GPflow 0.4.0 and tensorflow 1.2 on a Linux machine and a Mac. Note that latest GPflow breaks backward-compatibility.

We provide several test scripts (regression and classification) to demonstrate the usage. Running these examples should the results similar to the following figures:

Thang D. Bui, Cuong V. Nguyen and Richard E. Turner

@inproceedings{BuiNguTur17,
  title =  {Streaming sparse {G}aussian process approximations},
  author =   {Bui, Thang D. and Nguyen, Cuong V. and Turner, Richard E.},
  booktitle = {Advances in Neural Information Processing Systems 30},
  year =   {2017}
}

@article{BuiYanTur16,
  title={A Unifying Framework for Sparse {G}aussian Process Approximation using {P}ower {E}xpectation {P}ropagation},
  author={Thang D. Bui and Josiah Yan and Richard E. Turner},
  journal={arXiv preprint arXiv:1605.07066},
  year={2016}
}