Description:

The GPML toolbox implements approximate inference algorithms for Gaussian processes such as Expectation Propagation, the Laplace Approximation and Variational Bayes for a wide class of likelihood functions for both regression and classification. It comes with a big algebra of covariance, likelihood, mean and hyperprior functions allowing for flexible modeling. The code is fully compatible to Octave 3.2.x.

Changes to previous version:

Logdet-estimation functionality for grid-based approximate covariances

• Lanczos subspace estimation

• Chebyshef polynomial expansion

More generic infEP functionality

• dense computations and sparse approximations using the same code

• covering KL inference as a special cas of EP

New infKL function contributed by Emtiyaz Khan and Wu Lin

• Conjugate-Computation Variational Inference algorithm

• much more scalable than previous versions

Time-series covariance functions on the positive real line

• covW (i-times integrated) Wiener process covariance

• covOU (i-times integrated) Ornstein-Uhlenbeck process covariance (contributed by Juan Pablo Carbajal)

• covULL underdamped linear Langevin process covariance (contributed by Robert MacKay)

• covFBM Fractional Brownian motion covariance

New covariance functions

• covWarp implements k(w(x),w(z)) where w is a "warping" function

• covMatern has been extended to also accept non-integer distance parameters

BibTeX Entry: Download

Supported Operating Systems: Agnostic, Platform Independent

Data Formats: Matlab, Octave

Tags: Classification, Regression, Approximate Inference, Gaussian Processes

Archive: download here

Other available revisons

Version	Changelog	Date
4.1	Logdet-estimation functionality for grid-based approximate covariances Lanczos subspace estimation Chebyshef polynomial expansion More generic infEP functionality dense computations and sparse approximations using the same code covering KL inference as a special cas of EP New infKL function contributed by Emtiyaz Khan and Wu Lin Conjugate-Computation Variational Inference algorithm much more scalable than previous versions Time-series covariance functions on the positive real line covW (i-times integrated) Wiener process covariance covOU (i-times integrated) Ornstein-Uhlenbeck process covariance (contributed by Juan Pablo Carbajal) covULL underdamped linear Langevin process covariance (contributed by Robert MacKay) covFBM Fractional Brownian motion covariance New covariance functions covWarp implements k(w(x),w(z)) where w is a "warping" function covMatern has been extended to also accept non-integer distance parameters	November 27, 2017, 19:26:13
4.0	A major code restructuring effort did take place in the current release unifying certain inference functions and allowing more flexibility in covariance function composition. We also redesigned the whole derivative computation pipeline to strongly improve the overall runtime. We finally include grid-based covariance approximations natively. More generic sparse approximation using Power EP unified treatment of FITC approximation, variational approaches VFE and hybrids inducing input optimisation for all (compositions of) covariance functions dropping the previous limitation to a few standard examples infFITC is now covered by the more generic infGaussLik function Approximate covariance object unifying sparse approximations, grid-based approximations and exact covariance computations implementation in cov/apx, cov/apxGrid, cov/apxSparse generic infGaussLik unifies infExact, infFITC and infGrid generic infLaplace unifies infLaplace, infFITC_Laplace and infGrid_Laplace Hiearchical structure of covariance functions clear hierachical compositional implementation no more code duplication as present in covSEiso and covSEard pairs two mother covariance functions covDot for dot-product-based covariances and covMaha for Mahalanobis-distance-based covariances a variety of modifiers: eye, iso, ard, proj, fact, vlen more flexibility as more variants are available and possible all covariance functions offer derivatives w.r.t. inputs Faster derivative computations for mean and cov functions switched from partial derivatives to directional derivatives simpler and more concise interface of mean and cov functions much faster marginal likelihood derivative computations simpler and more compact code New mean functions new mean/meanWSPC (Weighted Sum of Projected Cosines or Random Kitchen Sink features) following a suggestion by William Herlands new mean/meanWarp for constructing a new mean from an existing one by means of a warping function adapted from William Herlands New optimizer added a new minimize_minfunc, contributed by Truong X. Nghiem New GLM link function added the twice logistic link function util/glm_invlink_logistic2 Smaller fixes two-fold speedup of util/elsympol used by covADD by Truong X. Nghiem bugfix in util/logphi as reported by John Darby	October 19, 2016, 10:15:05
3.6	added a new inference function infGrid_Laplace allowing to use non-Gaussian likelihoods for large grids fixed a bug due to Octave evaluating norm([]) to a tiny nonzero value, modified all lik/lik*.m functions reported by Philipp Richter small bugfixes in covGrid and infGrid bugfix in predictive variance of likNegBinom due to Seth Flaxman bugfix in infFITC_Laplace as suggested by Wu Lin bugfix in covPP{iso,ard}	July 6, 2015, 12:31:28
3.5	mechanism for specifying hyperparameter priors (together with Roman Garnett and José Vallet) new inference method inf/infGrid allowing efficient inference for data defined on a Cartesian grid (together with Andrew Wilson) new mean/cov functions for preference learning: meanPref/covPref new mean/cov functions for non-vectorial data: meanDiscrete/covDiscrete new piecewise constant nearest neighbor mean function: meanNN new mean functions being predictions from GPs: meanGP and meanGPexact new covariance function for standard additive noise: covEye new covariance function for factor analysis: covSEfact new covariance function with varying length scale : covSEvlen make covScale more general to scaling with a function instead of a scalar bugfix in covGabor* and covSM (due to Andrew Gordon Wilson) bugfix in lik/likBeta.m (suggested by Dali Wei) bugfix in solve_chol.c (due to Todd Small) bugfix in FITC inference mode (due to Joris Mooij) where the wrong mode for post.L was chosen when using infFITC and post.L being a diagonal matrix bugfix in infVB marginal likelihood for likLogistic with nonzero mean function (reported by James Lloyd) removed the combination likErf/infVB as it yields a bad posterior approximation and lacks theoretical justification Matlab and Octave compilation for L-BFGS-B v2.4 and the more recent L-BFGS-B v3.0 (contributed by José Vallet) smaller bugfixes in gp.m (due to Joris Mooij and Ernst Kloppenburg) bugfix in lik/likBeta.m (due to Dali Wei) updated use of logphi in lik/likErf bugfix in util/solve_chol.c where a typing issue occured on OS X (due to Todd Small) bugfix due to Bjørn Sand Jensen noticing that cov_deriv_sq_dist.m was missing in the distribution bugfix in infFITC_EP for ttau->inf (suggested by Ryan Turner)	December 8, 2014, 13:54:38
3.4	derivatives w.r.t. inducing points xu in infFITC, infFITC_Laplace, infFITC_EP so that one can treat the inducing points either as fixed given quantities or as additional hyperparameters new GLM likelihood likExp for inter-arrival time modeling new GLM likelihood likWeibull for extremal value regression new GLM likelihood likGumbel for extremal value regression new mean function meanPoly depending polynomially on the data infExact can deal safely with the zero noise variance limit support of GP warping through the new likelihood function likGaussWarp	November 11, 2013, 14:46:52
3.3	new generalised linear model likelihoods: gamma, beta, inverse Gaussian new ard/iso covariances: covPPard, covMaternard, covLINiso new spectral covariances: covSM, covGaboriso and covGaborard new meta covariance to turn an arbitrary stationary covariance into a periodic covariance one: covPERard, covPERiso new periodic covariance with zero DC component and correct scaling: covPeriodicNoDC, covCos new variational inference approximation based on direct KL minimisation: infKL improved inf/infVB double loop scheme so that only very few likelihood properties are required; infVB is now internally a sequence of infLaplace runs improved inf/infLaplace to be more generic so that optimisers other than scaled Newton can be used improved inf/infEP so that the internal variables (mu,Sigma) now represent the current posterior approximation	October 22, 2013, 15:34:05
3.2	We now support inference on large datasets using the FITC approximation for non-Gaussian likelihoods for EP and Laplace's approximation. New likelihood functions: mixture likelihood, Poisson likelihood, label noise. We added two MCMC samplers.	January 21, 2013, 15:34:50
3.1	We now support inference on large datasets using the FITC approximation by Ed Snelson. The covariance function interface had to be slightly modified.	September 28, 2010, 05:51:56
3.0	Initial Announcement on mloss.org.	July 23, 2010, 12:13:58

Comments

No one has posted any comments yet. Perhaps you'd like to be the first?

Leave a comment

You must be logged in to post comments.