PaperlayerView exact revision on arXiv

Source-linked AI summary

Gaussian Processes for Nonlinear Signal Processing

Fernando Pérez-Cruz, Steven Van Vaerenbergh, Juan José Murillo-Fuentes, Miguel Lázaro-Gredilla, Ignacio Santamaria

arXiv:1303.2823v2cs.LGcs.ITstat.ML

TL;DR

Signal processing has used Gaussian processes less widely than machine learning despite their suitability for nonlinear estimation and probabilistic prediction. This tutorial presents GP regression as a nonlinear Bayesian extension of MMSE/Wiener filtering, then develops scalable, adaptive, non-Gaussian, and classification extensions. It concludes that these extensions support uncertainty-aware and computationally practical signal-processing applications, including wireless communications.

  • Problem

    GPs are well established in machine learning but less widely used in signal processing because of computational demands, limited perceived flexibility, and unfamiliarity with their algorithmic role.

  • Method

    The tutorial formulates GP regression as a nonlinear Bayesian extension of MMSE/Wiener filtering and reviews sparse, adaptive, non-Gaussian, and classification extensions.

  • Results

    GPC provides more accurate posterior probabilities than an SVM-based equalizer and its BER is close to the optimal known-channel BCJR equalizer.

  • Takeaways & Limitations

    GPs provide uncertainty estimates, evidence-based hyper-parameter learning, scalable inference, and iterative adaptation for signal-processing applications.

Abstract

from arXiv · show

Gaussian processes (GPs) are versatile tools that have been successfully employed to solve nonlinear estimation problems in machine learning, but that are rarely used in signal processing. In this tutorial, we present GPs for regression as a natural nonlinear extension to optimal Wiener filtering. After establishing their basic formulation, we discuss several important aspects and extensions, including recursive and adaptive algorithms for dealing with non-stationarity, low-complexity solutions, non-Gaussian noise models and classification scenarios. Furthermore, we provide a selection of relevant applications to wireless digital communications.

I. INTRODUCTION

Gaussian processes are presented as Bayesian tools for nonlinear estimation, extending familiar MMSE and Wiener filtering ideas for signal-processing practitioners. The tutorial addresses their limited adoption by explaining their formulation, flexibility, computational issues, and relevant extensions.

  • Motivation: GPs provide Bayesian regression, classification, and dimensionality-reduction tools with full conditional statistical descriptions for predictions.Their probabilistic predictions can support confidence intervals and hyper-parameter setting.
  • Motivation: Signal-processing adoption of GPs has been limited by computational demands, unfamiliar classification intuition, perceived inflexibility, and their association primarily with noise models.These factors are contrasted with the wider use of neural networks and SVMs in signal processing.
  • Contribution: The tutorial introduces GPs as a natural nonlinear Bayesian extension of linear MMSE and Wiener filtering.It connects nonlinear least-squares solutions, flexible priors, evidence-based hyper-parameter setting, and full predictive uncertainty.
  • Scope: The paper surveys GP regression, recursive solutions, covariance design, computational simplification, non-stationarity, classification, and wireless-communication applications.The tutorial is organized into three parts covering formulation, key GP extensions, and signal-processing examples.
  • Problem: Nonlinear conditional mean estimation from finite samples can overfit or become suboptimal when linear-Gaussian assumptions do not hold.The passage identifies density-estimation convergence issues and restricted distributional models as sources of difficulty.

B. Gaussian Processes for Regression

Gaussian process regression models a latent nonlinear function with a Gaussian-process prior and conditions on observed training data to produce predictive distributions. Its mean matches kernel least-squares or kernel ridge regression, while its uncertainty and evidence-based hyper-parameter learning extend standard kernel methods.

  • Relation to MMSE: The latent-function formulation relaxes assumptions that inputs and outputs are linearly related or that the output distribution is Gaussian, while retaining a Gaussian conditional observation model.The tutorial notes that this makes GP a natural nonlinear extension to MMSE estimation.
  • Model formulation: GPR represents the input-output relation through a latent function drawn from a Gaussian-process prior characterized by a mean and covariance function.Finite input sets induce multivariate Gaussian distributions, enabling Bayesian conditioning on training labels.
  • Inference: Bayes’ rule, marginalization, and Gaussian conditioning produce the posterior predictive distribution for a new input.With Gaussian likelihoods, conjugacy makes the posterior analytically tractable; non-Gaussian observation models require extensions such as warped GPs.
  • Prediction: GPR provides a predictive mean and variance that include uncertainty from both the latent-function approximation and the observation model.The predictive mean is also the mean of p(y|x, Dn), while the variance accounts for observation noise.
  • Relation to kernel methods: The GPR mean equals kernel least-squares or kernel ridge regression for the same covariance matrix, but GPR additionally supplies error bars.Unlike typical SVM or KRR workflows, evidence sampling or maximization can fit many covariance hyper-parameters without grid-search cross-validation.

C. An example

The example illustrates how a GP posterior combines data-supported mean predictions with uncertainty that expands where observations are sparse. With 20 training points, the posterior is visualized through its mean, error bars, and sampled functions.

  • Setup: The example uses 20 training points, a fixed Gaussian kernel, and noise standard deviation σν = 0.1 to visualize the posterior.The plot shows the posterior mean, error bars µy ± 2σy, and five posterior samples.
  • Posterior behavior: Where observations are dense, GPR gives an accurate latent-function estimate with small error bars near the observation-noise level.The central region contains most of the data points.
  • Posterior behavior: Where observations are absent, the prediction returns toward the prior mean and variance, while sparse data widen uncertainty and constrain the mean near available samples.For x > 3, the posterior follows the zero-mean prior with ±2 uncertainty; the left region shows the mixed effect of sparse samples and the prior.

D. Recursive GPs

Online GPs update predictions recursively as new data arrive, avoiding recomputation of the full batch solution. The formulation also supports adaptation to non-stationary scenarios, while matrix growth remains a central practical limitation.

  • D. Recursive GPs: Online GPs recursively update predictions when each new datum arrives, reducing the need to recalculate the entire batch solution.The predicted distribution for the new output is computed first, followed by recursive updates using partitioned-matrix and Woodbury identities.
  • D. Recursive GPs: The recursive formulation updates posterior means and covariances for all available samples, making prediction changes easier to interpret.This formulation also makes adaptation to non-stationary scenarios straightforward.
  • D. Recursive GPs: The mean update applies a correction proportional to the estimation error, while only one of the related matrices needs to be stored and updated online.The relation between the covariance and inverse covariance matrices permits this storage choice.
  • D. Recursive GPs: The recursive GP mean update is equivalent to kernel recursive least-squares, but the GP additionally estimates posterior covariance and confidence intervals.KRLS implementations typically limit or fix matrix growth, whereas the GP formulation provides the full posterior distribution.
  • D. Recursive GPs: Unbounded matrix growth is the main limitation of the KRLS formulation, motivating implementations that limit growth or fix matrix sizes.The limitation appears in the recursively growing matrices used by the formulation.

E. Connection to MMSE: GPR with a linear latent function

A GP with a linear latent function recovers Bayesian linear estimation and connects kernel-based nonlinear estimation to MMSE filtering. The prior variance controls the relation to sampled MMSE, while the kernel formulation avoids explicitly specifying the nonlinear mapping.

  • E. Connection to MMSE: GPR with a linear latent function: Replacing the latent function with a linear model gives a spherical-Gaussian prior over its weights and a Gaussian posterior under Gaussian prior and likelihood.The posterior mean and covariance are obtained by standard Bayesian linear-model inference.
  • E. Connection to MMSE: GPR with a linear latent function: The Bayesian linear model's predictive distribution is Gaussian, with its mean and variance obtained from the posterior over the weights.The prediction for a general input is then expressed through the corresponding Bayesian quantities.
  • E. Connection to MMSE: GPR with a linear latent function: As the prior variance σ2_w tends to infinity, the Bayesian estimate becomes the sampled version of the MMSE solution.The prior's influence on the mean and covariance fades as more data become available.
  • E. Connection to MMSE: GPR with a linear latent function: Using the inner product between inputs scaled by the prior variance produces the kernel matrix representing the GP prior.The kernel matrix must include the prior width because it represents the Gaussian-process prior.
  • E. Connection to MMSE: GPR with a linear latent function: Substituting the Bayesian linear-model covariance expression recovers the GP predictive mean and variance without explicitly specifying the nonlinear mapping.This connects Bayesian linear estimation with nonlinear kernel or covariance-function estimation.

III. COVARIANCE FUNCTIONS

Covariance functions determine the flexibility and possible solutions of Gaussian-process regression, so their design and hyperparameter selection are central. The tutorial discusses composite kernels, evidence-based learning, and the computational cost of fully Bayesian hyperparameter integration.

  • III. COVARIANCE FUNCTIONS: The covariance function plays the role of a kernel, determining input relationships, how quickly functions change, and the possible GPR solutions.Designing an effective covariance function is crucial for accurate nonlinear solutions.
  • III. COVARIANCE FUNCTIONS: Covariance functions must be positive semidefinite and can be constructed by adding or multiplying simpler covariance matrices with positive hyperparameters.Kernel design should use problem information to obtain accurate solutions with as few samples as possible.
  • III. COVARIANCE FUNCTIONS: The example composite covariance combines an RBF term, a linear term, and a noise-variance term governed by hyperparameters θ.The RBF component uses separate input-dimension length-scales and provides a generic nonlinear regressor.
  • III. COVARIANCE FUNCTIONS: Additional covariance terms, including the Matérn kernel, can model faster transitions, extending the available function behavior.The noise variance can also be treated as an additional hyperparameter learned from data.
  • III. COVARIANCE FUNCTIONS: Fully Bayesian hyperparameter treatment requires sampling or approximation because the hyperparameter posterior is not available in closed form.MCMC may require several hundred to several thousand posterior samples, making the approach computationally intensive.
  • III. COVARIANCE FUNCTIONS: Maximum-likelihood hyperparameter selection is a faster alternative for time-sensitive applications, although its optimization problem is nonconvex.It is standard in the community but is not a purely Bayesian solution.

DATA SETS

Sparse GP methods reduce the cost of inference by approximating the covariance structure with a smaller active set or other structural simplifications. Their trade-offs include flexibility, convergence behavior, and the need for output transformations when target distributions are difficult to model.

  • DATA SETS: Exact GP inference requires covariance-matrix inversion with O(n3) cost, motivating sparse approximations intended to scale linearly with data size.These methods approximate the GP to make inference feasible for large databases.
  • DATA SETS: Sparse GP proposals can be unified as exact inference under modified priors that induce rank-m covariance matrices and reduce inference cost to O(m2n).Optional block-diagonal correction terms can accompany the low-rank covariance.
  • DATA SETS: SPGP selects active-set bases anywhere in the input space rather than restricting them to training samples, increasing the flexibility of the finite expansion.The selection is typically performed by evidence maximization.
  • DATA SETS: Increasing the number of SPGP bases does not generally converge to the full GP solution because its active set is unconstrained by the input data.This unconstrained selection can lead to overfitting in pathological cases.
  • DATA SETS: A variational sparse GP proposal guarantees convergence to the full GP solution while retaining an unconstrained active set.This addresses the convergence limitation identified for SPGP.
  • DATA SETS: Online sparsity commonly uses pruning to restrict the active set, and advanced pruning may improve efficiency at a fixed sparsity level.Other reduced-cost approaches use accelerated matrix-vector multiplication or compactly supported covariance functions.
  • DATA SETS: Output transformations such as logarithms or learned warping functions can improve GP modeling when target variables are difficult to represent directly.Warping parameters can be learned by maximizing the model evidence, with monotonic parametric or variational nonparametric options.

VI. TRACKING NON-STATIONARY SCENARIOS: LEARNING TO FORGET

The paper adapts online GP regression to non-stationary mappings by forgetting older samples, reducing their influence while retaining an inexpensive online algorithm. This forgetting operation is also exactly equivalent to inference in a spatio-temporal GP with an appropriate temporal covariance.

  • Forgetting in online GP regression: KRLS-T adds a forgetting step after each online update to reduce the influence of older samples.The posterior is shifted toward the prior for 0 < λ < 1, without needing to store or update C−1.
  • Spatio-temporal interpretation: The controlled forgetting procedure is principled rather than ad hoc: it corresponds exactly to a non-stationary GP model using an augmented input space with time stamps.The model uses a spatio-temporal covariance to account for changes in the latent mapping.
  • Spatio-temporal interpretation: With a temporal covariance parameterized by λ, spatio-temporal GP inference is exactly equivalent to applying the forgetting update after each online update.This equivalence provides an inexpensive online implementation for non-stationary inference.
  • Parameter interpretation: λ models the speed at which the mapping varies and can be selected principledly with Type-II ML rather than chosen only heuristically.With a linear spatial covariance, the model reduces to linear extended RLS filtering.
  • Illustration: For λ = 0.8, one forgetting update moves the predictive mean toward zero and expands the uncertainty region toward the prior variance of 1.The figure illustrates both the mean displacement and uncertainty expansion after forgetting.

A. Tracking a time-selective nonlinear communication channel

The tutorial evaluates KRLS-T on a nonlinear Rayleigh fading channel formed by a saturating nonlinearity followed by a time-varying linear channel. It compares simulated slow- and fast-fading cases with realistic test-bed measurements and reports comparisons against linear and kernel adaptive filters.

  • Channel model: The communication example uses a memoryless saturating nonlinearity followed by a time-varying linear channel.This model represents a channel such as one arising when a broadcast or satellite amplifier operates near saturation.
  • Simulation setup: The simulated channel has 5 randomly generated paths, y = tanh(x), and scenarios with fdT = 10−4 and fdT = 10−3 for slow and fast fading.The higher Doppler-frequency scenario changes faster and is therefore the more difficult tracking problem.
  • Comparative evaluation: KRLS-T is compared with NLMS, EX-RLS, and QKLMS using Gaussian kernels for the kernelized methods and bounded memory for KRLS-T.KRLS-T uses Type-II ML on a separate set of 500 test samples, with memory limited to 100 bases.
  • Simulation results: KRLS-T outperforms the other algorithms with a significant margin in both simulated fading scenarios.The comparison includes linear NLMS and EX-RLS alongside kernelized QKLMS.
  • Test-bed evaluation: Test-bed measurements use OFDM signals with the amplifier near saturation and fast time-varying channels around fdT = 10−3.The measured tracking results are described as similar to those of the simulated setup.

A. Probabilistic channel equalization

Probabilistic GPC equalization produces posterior estimates that better support channel decoding than SVM-based equalization, with BER close to the optimal solution. The tutorial also positions probabilistic uncertainty as a broader advantage of Gaussian-process methods.

  • Probabilistic outputs: When probabilistic outputs matter, GPC retains similar predictive performance to SVMs while naturally incorporating confidence intervals into its predictions.This probabilistic output is important because channel decoders operate optimally when given accurate posterior estimates for each symbol.
  • Posterior calibration: GPC predictions are significantly closer to the perfect-calibration diagonal than SVM predictions, providing more accurate posterior probabilities to the channel decoder.The comparison uses a highly noisy scenario with normalized signal-to-noise ratio of 2 dB.
  • Decoding performance: GPC-based equalization clearly outperforms SVM-based equalization in BER after probabilistic channel decoding and approaches the optimal known-channel BCJR solution.The optimal reference uses the known channel and a forward-backward BCJR equalizer.
  • Uncertainty estimation: GPR provides uncertainty estimates that account for both approximation error and likelihood-model error, whereas KLS assumes likelihood-defined error bars that are constant across the input space.This gives GPR input-dependent uncertainty information alongside its mean estimate.
  • Extensions: The tutorial extends GP methods to non-Gaussian noise, classification, reduced-complexity computation, and iterative adaptation for non-stationary environments.The stated computational reduction is from cubic to linear in the number of training examples without significantly affecting mean and error-bar predictions.
Loading 1303.2823v2…