AugmentedGaussianProcesses.jl is a Julia package in development for **Data Augmented Sparse Gaussian Processes**. It contains a collection of models for different **gaussian and non-gaussian likelihoods**, which are transformed via data augmentation into **conditionally conjugate likelihood** allowing for **extremely fast inference** via block coordinate updates. There are also more options to use more traditional **variational inference** via quadrature or Monte Carlo integration.

**BayesianSVM**: A Classifier with a likelihood equivalent to the classic SVM IJulia example/Reference**Logistic**: A Classifier with a Bernoulli likelihood with the logistic link IJulia example/Reference

**Gaussian**: The standard Gaussian Process regression model with a Gaussian Likelihood (no data augmentation was needed here) IJulia example/Reference**StudentT**: The standard Gaussian Process regression with a Student-t likelihood (the degree of freedom ν is not optimizable for the moment) IJulia example/Reference**Laplace**: Gaussian Process regression with a Laplace likelihood IJulia example/(No reference at the moment)**Heteroscedastic**: Regression with non-stationary noise, given by an additional GP. (no reference at the moment)

**Discrete Poisson Process**: Estimating a the Poisson parameter λ at every point (as λ₀σ(f)). (no reference at the moment)**Negative Binomial**: Estimating the success probability at every point for a negative binomial distribution (no reference at the miment)

**Logistic-SoftMax**: A modified version of the softmax where the exponential is replaced by the logistic function IJulia example/Reference

## Multi-Ouput models

- It is also possible to create a multi-ouput model where the outputs are a linear combination of inducing variables see IJulia example in preparation/[Reference][neuripsmultiouput]

**Probit**: A Classifier with a Bernoulli likelihood with the probit link**Online**: Allowing for all algorithms to work online as well

The package requires at least Julia 1.1
Run `julia`

, press `]`

and type `add AugmentedGaussianProcesses`

, it will install the package and all its dependencies.

A complete documentation is available in the docs. For a short start now you can use this very basic example where `X_train`

is a matrix `N x D`

where `N`

is the number of training points and `D`

is the number of dimensions and `Y_train`

is a vector of outputs (or matrix of independent outputs).

```
using AugmentedGaussianProcesses;
using KernelFunctions
model = SVGP(X_train,Y_train,SqExponentialKernel(1.0),LogisticLikelihood(),AnalyticSVI(100),64)
train!(model,100)
Y_predic = predict_y(model,X_test) #For getting the label directly
Y_predic_prob, Y_predic_prob_var = proba_y(model,X_test) #For getting the likelihood (and likelihood uncertainty) of predicting class 1
```

Both documentation and examples/tutorials are available.

Check out my website for more news

"Gaussian Processes for Machine Learning" by Carl Edward Rasmussen and Christopher K.I. Williams

UAI 19' "Multi-Class Gaussian Process Classification Made Conjugate: Efficient Inference via Data Augmentation" by Théo Galy-Fajou, Florian Wenzel, Christian Donner and Manfred Opper https://arxiv.org/abs/1905.09670

ECML 17' "Bayesian Nonlinear Support Vector Machines for Big Data" by Florian Wenzel, Théo Galy-Fajou, Matthäus Deutsch and Marius Kloft. https://arxiv.org/abs/1707.05532

AAAI 19' "Efficient Gaussian Process Classification using Polya-Gamma Variables" by Florian Wenzel, Théo Galy-Fajou, Christian Donner, Marius Kloft and Manfred Opper. https://arxiv.org/abs/1802.06383

NeurIPS 18' "Moreno-Muñoz, Pablo, Antonio Artés, and Mauricio Álvarez. "Heterogeneous multi-output Gaussian process prediction." Advances in Neural Information Processing Systems. 2018." [https://papers.nips.cc/paper/7905-heterogeneous-multi-output-gaussian-process-prediction][neuripsmultiouput]

UAI 13' "Gaussian Process for Big Data" by James Hensman, Nicolo Fusi and Neil D. Lawrence https://arxiv.org/abs/1309.6835

JMLR 11' "Robust Gaussian process regression with a Student-t likelihood." by Jylänki Pasi, Jarno Vanhatalo, and Aki Vehtari. http://www.jmlr.org/papers/v12/jylanki11a.html

01/25/2018

about 6 hours ago

619 commits