VariationalGaussianProcess<T>

Models & Types · AiDotNet.GaussianProcesses

Implements a Variational Gaussian Process (VGP) using variational inference for exact data.

Inheritance: ModelBase<T, Matrix<T>, Vector<T>> → GaussianProcessBase<T> → VariationalGaussianProcess<T>

Implements: IGeneratedParameterRegistrar<T>

For Beginners

The Variational Gaussian Process (VGP) is a probabilistic model that uses variational inference to approximate the posterior distribution over functions.

Unlike SVGP which uses inducing points for scalability, VGP works with all training points but uses variational inference to handle non-Gaussian likelihoods (like for classification or robust regression with non-Gaussian noise).

Key differences from standard GP:

  • Standard GP: Assumes Gaussian likelihood, has closed-form solution
  • VGP: Can handle any likelihood, uses optimization to find approximate posterior

When to use VGP:

  • When you have non-Gaussian likelihoods (classification, count data, etc.)
  • When you want uncertainty quantification with flexible likelihood models
  • When your dataset is small enough to use all points (up to ~5000 points)

For large datasets with Gaussian likelihood, use SparseVariationalGaussianProcess instead.

How It Works

Implementation Note: The ELBO (Evidence Lower Bound) computation uses a simplified trace approximation (S ≈ K) for numerical stability. This approximation may affect the absolute ELBO values, making them unreliable for cross-model comparison. For hyperparameter optimization, consider using the prediction accuracy on a held-out validation set instead.

Constructors

Constructor Summary
VariationalGaussianProcess(IKernelFunction<T> kernel, VGPLikelihood likelihood, double noiseVariance, double learningRate, int maxIterations, double tolerance, MatrixDecompositionType decompositionType) Initializes a new instance of the VariationalGaussianProcess class.

Methods

Method Summary
T ComputeELBO() Computes the Evidence Lower Bound (ELBO) for the current variational approximation.
ValueTuple<T, T> Predict(Vector<T> x)
void Fit(Matrix<T> X, Vector<T> y) Trains the VGP model using variational inference.
void UpdateKernel(IKernelFunction<T> kernel)