InducingPointKernel<T>

Models & Types · AiDotNet.Kernels

Inducing Point Kernel for sparse Gaussian Process approximations.

Implements: IKernelFunction<T>

For Beginners

The Inducing Point Kernel is fundamental to making GPs scale to large datasets. Instead of working with all N training points (O(N³) complexity), we use a smaller set of M “inducing points” (O(NM² + M³) complexity, where M << N).

The key idea: Instead of modeling the full function f, we model:

  1. The function values at M inducing points: u = f(Z)
  2. The conditional distribution: f(x) | u

The approximation makes f(x) conditionally independent given the inducing values u. The closer the inducing points are to your data, the better the approximation.

This kernel computes the “Q_ff” approximation: Q(x, x’) = k(x, Z) × k(Z, Z)⁻¹ × k(Z, x’)

Where:

  • Z are the inducing point locations (M × d)
  • k(x, Z) is the cross-covariance (1 × M)
  • k(Z, Z) is the inducing point covariance (M × M)

This is the Nyström approximation of the full kernel matrix.

How It Works

Usage patterns:

  • Use with SparseVariationalGaussianProcess for scalable GP regression
  • Inducing points can be learned via optimization
  • Good initialization: subset of training data or k-means centroids

Constructors

Constructor Summary
InducingPointKernel(IKernelFunction<T> baseKernel, Matrix<T> inducingPoints, double jitter) Initializes an Inducing Point Kernel.

Properties

Property Summary
IKernelFunction<T> BaseKernel { get; } Gets the base kernel.
int NumInducingPoints { get; } Gets the number of inducing points.

Methods

Method Summary
Matrix<T> ComputeCrossCovariance(Matrix<T> X) Computes the cross-covariance between test points and inducing points.
Matrix<T> GetCholeskyFactor() Gets the Cholesky factor of k(Z, Z).
Matrix<T> GetInducingPointCovariance() Gets the inducing point covariance matrix k(Z, Z).
Matrix<T> GetInducingPoints() Gets a copy of the inducing points.
T Calculate(Vector<T> x1, Vector<T> x2) Calculates the approximate kernel value (Nyström approximation).
T EstimateApproximationError(Matrix<T> X) Estimates the approximation quality at a set of test points.
void PrecomputeMatrices() Precomputes the inducing point matrices for efficiency.
void UpdateInducingPoints(Matrix<T> newInducingPoints) Updates the inducing point locations.