[ANN] AugmentedGaussianProcesses.jl

Just want to make so shameless advertisement for my Gaussian Process package : AugmentedGaussianProcesses.jl
With it you can work efficiently with non-gaussian likelihood (as well as gaussian of course) while scaling with high number of data.
It features :

  • Non-Gaussian likelihoods : Student-T, Bernoulli (logistic link), Bayesian-SVM (hinge loss) and a multi-class likelihood (similar to softmax), and more incoming
  • Scalability via inducing points (scalable to 1e6 points via stochastic updates)
  • Inference via : Numerical Variational Inference (gradients approximated numerically), Analytic Variational Inference (via an augmentation trick!), Gibbs Sampling
  • Hyperparameter optimization included (but to be improved)

Feature incoming :

  • Online learning
  • More kernel functions
  • Integration of AD

Please check it out, any feedback is welcome :slight_smile:
PS: Also I am alone on this project so if anyone is interested to work on this as well, this would be amazing


This is a fantastic package! Nice work

Thanks! I appreciate!
Are you using GP for your work? Tell me if there are any feature you are missing :slight_smile:

1 Like

Hi Theogf! yes i’m using it to perform regression on the results of a very noise MCMC simulation, perhaps some more examples on how different kernel produce different results, and the ability to load a trained GP from file.
Anyway it is a very easy to use and powerful package, thank you again for your work!