┌───────────────────┐
│⢰⢦⣀⣀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⡆⠀⠀⢀⣴⣶⢄⠀⠀⠀⠀⠀⠀⠀⠀│
│⢨⣿⣷⣻⣵⡀⠀⠀⠀⠀⠀⠀⠀⠀⠀⡇⠀⣠⣾⣿⣿⣿⡆⠀⠀⠀⠀⠀⠀⠀│
│⠨⡿⣿⣿⣿⣧⠀⠀⠀⠀⠀⠀⠀⠀⠀⡇⣼⣿⣿⠿⣿⣾⣿⡀⠀⠀⠀⠀⠀⠀│
│⠸⠋⠑⢿⣿⣿⡆⠀⠀⠀⠀⠀⠀⠀⠀⣿⣿⣿⡟⠉⠺⣻⣿⣧⠀⠀⠀⠀⠀⠀│
│⠀⠀⠀⠀⠹⣿⣿⡀⠀⠀⠀⠀⠀⠀⣼⣿⣿⠝⠀⠀⠀⠈⢿⣿⡆⠀⠀⠀⠀⠀│
│⠉⠉⠉⠉⠉⢻⣿⣯⠉⠉⠉⠉⠉⣹⣿⣿⠏⠉⠉⠉⠉⠉⠉⣿⣿⣍⡉⠉⠉⠉│
│⠀⠀⠀⠀⠀⠀⢿⣿⡆⠀⠀⢀⣮⣿⣿⡏⠀⠀⠀⠀⠀⠀⠀⠸⣿⣏⢟⡄⡜⠀│
│⠀⠀⠀⠀⠀⠀⠈⢿⣿⣦⣶⣫⣿⣿⡏⡇⠀⠀⠀⠀⠀⠀⠀⠀⠹⣿⣯⣫⣦⠂│
│⠀⠀⠀⠀⠀⠀⠀⠘⢿⣿⣿⣿⡿⡟⠀⡇⠀⠀⠀⠀⠀⠀⠀⠀⠀⠹⣿⣿⣿⡄│
│⠀⠀⠀⠀⠀⠀⠀⠀⠀⠙⠾⠿⠏⠀⠀⡇⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠹⢿⠿⠀│
│⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⡇⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠱⠀│
└───────────────────┘
This package implements sparse Gaussian processes models using doubly stochastic variational inference.
Unlike essentially all other Gaussian process packages, SparseGaussianProcesses.jl does not work with means and covariances. Instead, it uses the path-wise sampling technique to implement entire function draws from Gaussian process posteriors, which can be evaluated deterministically at arbitrary locations once sampled.
It supports models of the form
(f | u)(.) = (Ag)(.) + K_{(.)z} (K_{zz} + \Lambda)^{-1} (u - (Bg)(z) - \epsilon)
where g ~ GP(0, k), u ~ N(\mu, \Sigma), \epsilon ~ N(0, \Lambda), and A, B are inter-domain operators such as the identity, gradient, or convolutional patch map.
This little-known formula defines a Gaussian process with precisely the correct mean and variance of a standard sparse Gaussian process.
The following features are planned for this package.
- Evaluation of entire function draws at arbitrary locations.
- Posterior sample paths are fully differentiable, assuming a sufficiently smooth kernel.
- Strong inter-domain support, including gradient and convolutional Gaussian processes.
- Fully supports training on GPU.
- Strong support for vector-valued processes.
- Strong support for non-Euclidean domains.
A set of examples are available in the examples/ folder.
This package is under development, and contributions are welcome! The best way to get in touch regarding development is on the Julia slack.
@article{wilson20,
Author = {James T. Wilson and Viacheslav Borovitskiy and Alexander Terenin and Peter Mostowski and Marc Peter Deisenroth},
Journal = {arXiv:2002.09309},
Title = {Efficiently sampling functions from Gaussian process posteriors},
Year = {2020}}