GPs at Large Scale¶
Any literature related to scaling GPs to astronomical size. 1 million points and higher. Also things with respectable orders of magnitude.
Fast Direct Methods for Gaussian Processes by Ambikasaran et. al. (2014)
Scales GPs with dimensions greater than 3.
-> Paper
-> Code
Constant-Time Predictive Distributions for GPs - Pleiss et. al. (2018)
Using MVM techniquees, they are able to make constant time predictive mean and variance estimates; addressing the computational bottleneck of predictive distributions for GPs.
-> Paper
-> Code
When Gaussian Process Meets Big Data: A Review of Scalable GPs - (2019)
A great review paper on GPs and how to scale them. Goes over most of the SOTA.
-> Paper
Exact GP on a Million Data Points - Wang et. al. (2019)
The authors manage to train a GP with multiple GPUs on a million or so data points. They use the matrix-vector-multiplication strategy. Quite amazing actually...
Randomly Projected Additive Gaussian Processes for Regression - Delbridge et. al. (12-2019)
-> Paper
-> Code (PyTorch)
Efficiently Sampling Functions from Gaussian Process Posteriors - Wilson et. al. (16-2020)
Uses a path-wise sampling scheme to efficiently sample for GP posteriors. Motivates the use for GP priors for monte carlo estimation schemes.
-> Blog
-> Video
-> Paper
Fast Matrix Square Roots with Applications to Gaussian Processes and Bayesian Optimization by Pleiss et. al. (2020)
-> Paper
-> Code
Sparse Cholesky factorization by Kullback-Leibler minimization - Schafer et, al. (2020)
-> Paper