Non-Stationary Spectral Kernels
نویسندگان
چکیده
We propose non-stationary spectral kernels for Gaussian process regression. Wepropose to model the spectral density of a non-stationary kernel function as amixture of input-dependent Gaussian process frequency density surfaces. Wesolve the generalised Fourier transform with such a model, and present a familyof non-stationary and non-monotonic kernels that can learn input-dependent andpotentially long-range, non-monotonic covariances between inputs. We deriveefficient inference using model whitening and marginalized posterior, and showwith case studies that these kernels are necessary when modelling even rathersimple time series, image or geospatial data with non-stationary characteristics.
منابع مشابه
Classes of Kernels for Machine Learning: A Statistics Perspective
In this paper, we present classes of kernels for machine learning from a statistics perspective. Indeed, kernels are positive definite functions and thus also covariances. After discussing key properties of kernels, as well as a new formula to construct kernels, we present several important classes of kernels: anisotropic stationary kernels, isotropic stationary kernels, compactly supported ker...
متن کاملNon-stationary Wiener filtering based on evolutionary spectral theory
In this paper we consider solutions to the non-stationary Wiener ltering problem using the evolutionary spectral theory. Two cases of interest result from the uncorrelation between the desired signal and the noise. One constrains the support of the generating kernels of the signals and the other imposes orthogonality on their innovation processes. The latter condition is more general and our so...
متن کاملGaussian Process Covariance Kernels for Pattern Discovery and Extrapolation
Gaussian processes are rich distributions over functions, which provide a Bayesian nonparametric approach to smoothing and interpolation. We introduce simple closed form kernels that can be used with Gaussian processes to discover patterns and enable extrapolation. These kernels are derived by modelling a spectral density – the Fourier transform of a kernel – with a Gaussian mixture. The propos...
متن کاملGaussian Process Kernels for Pattern Discovery and Extrapolation
Gaussian processes are rich distributions over functions, which provide a Bayesian nonparametric approach to smoothing and interpolation. We introduce simple closed form kernels that can be used with Gaussian processes to discover patterns and enable extrapolation. These kernels are derived by modelling a spectral density – the Fourier transform of a kernel – with a Gaussian mixture. The propos...
متن کاملSynchrosqueezing-based Transform and its Application in Seismic Data Analysis
Seismic waves are non-stationary due to its propagation through the earth. Time-frequency transforms are suitable tools for analyzing non-stationary seismic signals. Spectral decomposition can reveal the non-stationary characteristics which cannot be easily observed in the time or frequency representation alone. Various types of spectral decomposition methods have been introduced by some resear...
متن کامل