Click or drag to resize
SparseGP Class
Microsoft Research
A Gaussian Process distribution over functions, represented by a GP prior times a set of regression likelihoods on basis points.
Inheritance Hierarchy
SystemObject
  MicrosoftResearch.Infer.DistributionsSparseGP

Namespace: MicrosoftResearch.Infer.Distributions
Assembly: Infer.Runtime (in Infer.Runtime.dll) Version: 2.6.41128.1 (2.6.41128.1)
Syntax
[SerializableAttribute]
[Quality(QualityBand.Preview)]
public class SparseGP : IGaussianProcess, IDistribution<IFunction>, 
	ICloneable, HasPoint<IFunction>, Diffable, SettableToUniform, 
	CanGetLogProb<IFunction>, SettableTo<SparseGP>, SettableToProduct<SparseGP>, 
	SettableToProduct<SparseGP, SparseGP>, SettableToRatio<SparseGP>, 
	SettableToRatio<SparseGP, SparseGP>, SettableToPower<SparseGP>, 
	SettableToWeightedSum<SparseGP>, CanGetLogAverageOf<SparseGP>, CanGetLogAverageOfPower<SparseGP>, 
	CanGetAverageLog<SparseGP>, CanGetMean<IFunction>, Sampleable<IFunction>

The SparseGP type exposes the following members.

Constructors
  NameDescription
Public methodSparseGP(SparseGP)
Copy constructor
Public methodSparseGP(SparseGPFixed)
Constructs sparse GP, given basis etc
Public methodSparseGP(SparseGPFixed, Boolean)
Public methodSparseGP(SparseGPFixed, Boolean, VectorGaussian, IFunction)
Constructor from full specification
Top
Methods
  NameDescription
Public methodClearCachedValues
Function to signal recalculation of calculated parameters. This is called automatically if the fixed parameter class is swapped out, or if the kernel is changed, or if parameters are changed. It should also be called by any external program modifies the kernel or other fixed parameters in place
Public methodClone
Clone. Note that the fixed parameters and the rank1 list are just referenced
Public methodCovariance(IListVector)
Predictive coariance at a given list of points
Public methodCovariance(Vector, Vector)
Predictive covariance at a given pair of points
Public methodEquals
Public methodEvaluateMean
Evaluates the mean function of the GP
Public methodGetAverageLog
The expected logarithm of that distribution under this distribution
Public methodGetHashCode
Public methodGetLogAverageOf
Gets the log of the integral of the product of this SparseGP and that SparseGP
Public methodGetLogAverageOfPower
Get the integral of this distribution times another distribution raised to a power.
Public methodGetLogProb
Gets the log density for a given value
Public methodGetMean
Gets the mean function for the Sparse GP
Public methodIsUniform
Asks the distribution whether it is uniform
Public methodJoint
Predictive distribution at a given list of points
Public methodMarginal
Predictive distribution at a given point
Public methodMaxDiff
Max difference between two sparse GPs - used for convergence testing
Public methodMean(Vector)
Mean at a given point
Public methodMean(IListVector)
Mean at a given list of points
Public methodStatic memberPointMass
Creates a sparse GP point mass - i.e. all the mass is at a given function
Public methodSample
Samples from the Sparse Gaussian distribution This is only implemented for a 1-dimensional input space, and returns a simple linear spline function
Public methodSample(IFunction)
Samples from the Sparse Gaussian distribution This is only implemented for a 1-dimensional input space, and returns a simple linear spline function. result is ignored This argument is ignored
Public methodSetTo
Sets one sparse GP to another. Everything is copied except the FixedParameters and the lsit of rank 1 potentials which are referenced.
Public methodSetToPower
Sets this sparse GP the the power of another sparse GP
Public methodSetToProduct
Sets this instance to the product of two sparse GPs.
Public methodSetToRatio
Sets this instance to the ratio of two sparse GPs.
Public methodSetToSum
Sets this SparseGP distribution to the weighted sum of two other such distributions
Public methodSetToUniform
Sets to uniform
Public methodStatic memberUniform
Creates a uniform sparse GP
Public methodVariance
Predictive Variance at a given point
Top
Operators
  NameDescription
Public operatorStatic memberDivision
Creates a new SparseGP which the ratio of two other SparseGPs
Public operatorStatic memberMultiply
Creates a new SparseGP which the product of two other SparseGPs
Top
Fields
  NameDescription
Public fieldIncludePrior
Whether this sparse GP includes the prior
Public fieldInducingDist
The regression likelihoods that modify the prior.
Top
Properties
  NameDescription
Public propertyAlpha
Alpha - along with beta, this encodes the posterior means and covariances of the Sparse GP
Public propertyBeta
Beta - along with alpha, this encodes the posterior means and covariances of the Sparse GP
Public propertyFixedParameters
Sets and gets the fixed sparse parameters - parameters which are not changed by inference
Public propertyIsPointMass
Asks the distribution whether it is a point mass
Public propertyMean_B
m(B). This is a calculated Vector maintained by the class
Public propertyPoint
Sets or Gets a point. If not a point function, the get returns the mean function of the sparse GP
Public propertyVar_B_B
var(B, B). This is a calculated matrix maintained by the class
Top
Remarks

This distribution family comes from the paper "Sparse-posterior Gaussian Processes for general likelihoods" by Qi et al (2010), http://event.cwi.nl/uai2010/papers/UAI2010_0283.pdf

The state of the distribution is represented by (FixedParameters, IncludePrior, InducingDist, pointFunc). The GP prior and basis point locations are stored in FixedParameters. The regression likelihoods are stored as a single VectorGaussian called InducingDist. IncludePrior=false does not include the prior in the distribution (i.e. the distribution is degenerate). If pointFunc != null, the distribution is a point mass. If InducingDist is uniform and IncludePrior is false, the distribution is uniform. The GP prior is assumed to be non-uniform.

See Also