Skip to content

Overview

We will do a quick overview to show how we can account for input errors in Gaussian process regression models.

Problem Statement

Standard

y=f(x)+ϵy

where ϵyN(0,σy2). Let x=μx+Σx.

Observe Noisy Estimates

y=f(x)+ϵy

Observation Means only

y=f(μx)+ϵy

Posterior Predictions

μ(x)=k(K+λIN)1y=kα
νGP2=σy2+kk(K+σy2IN)1k

Linearized Approximation

Where we take the Taylor expansion of the predictive mean and variance function. The mean function stays the same:

μeGP(x)=kα

but the predictive variance term gets changed slightly:

νeGP2=νGP2(x)+μΣxμ+Tr{2νGP2(x)xx|x=μxΣx}

with the term in red being the derivative of the predictive mean function multiplied by the variance.

Notes: * Assumes known variance * Assumes D×D covariance matrix for multidimensional data * Quite inexpensive to implement * The 3rd term (the 2nd order component of the Taylor expansion) has been show to not make a huge difference

egp_moment1 = jax.jfwd(posterior, args_num=(None, 0))
egp_moment2 = jax.hessian(posterior, args_num=(None, 0))

Moment-Matching

Mean Predictions

μeGP(x)=qα

where:

qi=|Λ1Σx+I|1/2exp[12(μxi)(Σx+Λ)1(μxj)]

Variance Predictions

νeGP2

Variational

Assumes we have a variational distribution function

L(θ)=DKL[q(f)q(X)||p(f|X)p(X)]

Other Resources

Gaussian Process Model Zoo


Datasets

We use some toy datasets which including:

  1. "near square sine wave"
f(x)=sin(πccos(5+x2))
  1. The sigmoid curve
f(x)=11+exp(x)
  1. Mauna Loa Ice Core Data | Data Portal

$$

$$

  1. Spatial IASI Data