Gaussian process

About: Gaussian process is a research topic. Over the lifetime, 18944 publications have been published within this topic receiving 486645 citations. The topic is also known as: Gaussian stochastic process.

Characterisation of radar clutter as a spherically invariant random process

Abstract: A statistical characterisation of clutter as a complex random process is needed in the design of optimum detection schemes. The paper considers modelling complex clutter as a spherically invariant random process (SIRP), namely assuming that its PDFs can be expressed as non-negative definite quadratic forms, a generalisation of a Gaussian process. Relevant properties of SIRPs are summarised, and shown to comply with basic requirements such as circular symmetry of the joint PDF of the in-quadrature components or, equivalently, the uniformity of the phase distribution. A constraint of admissibility must be imposed on the envelope distribution, but most commonly used envelope distributions, including Weibull, contaminated Rayleigh and K-distribution are shown to be admissible. Although a general SIRP is not ergodic, a characterisation of the clutter process as an SIRP scanned in the ensemble is finally proposed, which restores ergodicity. The interpretation of this model in the light of already proposed composite scattering models is also discussed.

Sequential design of computer experiments for the estimation of a probability of failure

Abstract: This paper deals with the problem of estimating the volume of the excursion set of a function f:? d ?? above a given threshold, under a probability measure on ? d that is assumed to be known. In the industrial world, this corresponds to the problem of estimating a probability of failure of a system. When only an expensive-to-simulate model of the system is available, the budget for simulations is usually severely limited and therefore classical Monte Carlo methods ought to be avoided. One of the main contributions of this article is to derive SUR (stepwise uncertainty reduction) strategies from a Bayesian formulation of the problem of estimating a probability of failure. These sequential strategies use a Gaussian process model of f and aim at performing evaluations of f as efficiently as possible to infer the value of the probability of failure. We compare these strategies to other strategies also based on a Gaussian process model for estimating a probability of failure.

Optimal adaptive estimation: Structure and parameter adaption

Abstract: Optimal structure and parameter adaptive estimators have been obtained for continuous as well as discrete data Gaussian process models with linear dynamics. Specifically, the essentially nonlinear adaptive estimators are shown to be decomposable (partition theorem) into two parts, a linear nonadaptive part consisting of a bank of Kalman-Bucy filters and a nonlinear part that incorporates the adaptive nature of the estimator. The conditional-error-covariance matrix of the estimator is also obtained in a form suitable for on-line performance evaluation. The adaptive estimators are applied to the problem of state estimation with non-Gaussian initial state, to estimation under measurement uncertainty (joint detection-estimation) as well as to system identification. Examples are given of the application of the adaptive estimators to structure and parameter adaptation indicating their applicability to engineering problems.

Bayesian Deep Learning and a Probabilistic Perspective of Generalization

Abstract: The key distinguishing property of a Bayesian approach is marginalization, rather than using a single setting of weights. Bayesian marginalization can particularly improve the accuracy and calibration of modern deep neural networks, which are typically underspecified by the data, and can represent many compelling but different solutions. We show that deep ensembles provide an effective mechanism for approximate Bayesian marginalization, and propose a related approach that further improves the predictive distribution by marginalizing within basins of attraction, without significant overhead. We also investigate the prior over functions implied by a vague distribution over neural network weights, explaining the generalization properties of such models from a probabilistic perspective. From this perspective, we explain results that have been presented as mysterious and distinct to neural network generalization, such as the ability to fit images with random labels, and show that these results can be reproduced with Gaussian processes. We also show that Bayesian model averaging alleviates double descent, resulting in monotonic performance improvements with increased flexibility. Finally, we provide a Bayesian perspective on tempering for calibrating predictive distributions.

A Gaussian process framework for modelling instrumental systematics: application to transmission spectroscopy

Abstract: Transmission spectroscopy, which consists of measuring the wavelength-dependent absorption of starlight by a planet’s atmosphere during a transit, is a powerful probe of atmospheric composition. However, the expected signal is typically orders of magnitude smaller than instrumental systematics and the results are crucially dependent on the treatment of the latter. In this paper, we propose a new method to infer transit parameters in the presence of systematic noise using Gaussian processes, a technique widely used in the machine learning community for Bayesian regression and classification problems. Our method makes use of auxiliary information about the state of the instrument, but does so in a non-parametric manner, without imposing a specific dependence of the systematics on the instrumental parameters, and naturally allows for the correlated nature of the noise. We give an example application of the method to archival NICMOS transmission spectroscopy of the hot Jupiter HD 189733, which goes some way towards reconciling the controversy surrounding this data set in the literature. Finally, we provide an appendix giving a general introduction to Gaussian processes for regression, in order to encourage their application to a wider range of problems.