wzh1895 / GPFADS

Gaussian Process Factor Analysis with Dynamical Structure

GPFADS: Gaussian Process Factor Analysis with Dynamical Structure

Summary

This is the repository associated to: Non-reversible Gaussian processes for identifying latent dynamical structure in neural data - NeurIPS (2020)

The repo contains three main classes:

• a Kernel class allowing one to build various d-dimensional non-reversible multi-output kernels
• a GPregression class to run GP regression on input data
• a GPFA class to run GPFA with reversible or non-reversible priors

Requirements

• python-3
• scipy
• autograd

Non-time reversible multi-output GP kernels

Class for building non-reversible kernels

Basic demo

Example of how to build and sample from a non-time-reversible kernel

     python ./examples/draw_samples.py

Using kernel class build N-d multi-output GP non-rev GP.
Here covariances shown for N = 2 and with squared exponentials marginals and sample draws. Other marginals are available in class.

GP regression with non-reversible priors

Basic class to run GP-regression using non-rev prior

Basic demo

Run GP regression on toy data, optimise and show posterior mean inferred on test data whilst only conditioning on one of the two observed variables.

     python ./examples/gpregression_demo.py

GPFADS - Gaussian Process Factor Analysis with Dynamical Structure

a.k.a.: GPFA with non-reversible priors

Basic demo

Run GPFADS to unmix 2 dynamical processes and infer posterior mean on test data.

References

• Non-reversible Gaussian processes for identifying latent dynamical structure in neural data - NeurIPS (2020)