Learning multidimensional Fourier series with tensor trains

Sander Wahls, Visa Koivunen, H. Vincent Poor, Michel Verhaegen

Research output: Chapter in Book/Report/Conference proceedingConference contribution

3 Scopus citations

Abstract

How to learn a function from observations of inputs and noisy outputs is a fundamental problem in machine learning. Often, an approximation of the desired function is found by minimizing a risk functional over some function space. The space of candidate functions should contain good approximations of the true function, but it should also be such that the minimization of the risk functional is computationally feasible. In this paper, finite multidimensional Fourier series are used as candidate functions. Their impressive approximative capabilities are illustrated by showing that Gaussian-kernel estimators can be approximated arbitrarily well over any compact set of bandwidths with a fixed number of Fourier coefficients. However, the solution of the associated risk minimization problem is computationally feasible only if the dimension d of the inputs is small because the number of required Fourier coefficients grows exponentially with d. This problem is addressed by using the tensor train format to model the tensor of Fourier coefficients under a low-rank constraint. An algorithm for least-squares regression is derived and the potential of this approach is illustrated in numerical experiments. The computational complexity of the algorithm grows only linearly both with the number of observations N and the input dimension d, making it feasible also for large-scale problems.

Original languageEnglish (US)
Title of host publication2014 IEEE Global Conference on Signal and Information Processing, GlobalSIP 2014
PublisherInstitute of Electrical and Electronics Engineers Inc.
Pages394-398
Number of pages5
ISBN (Electronic)9781479970889
DOIs
StatePublished - Feb 5 2014
Event2014 IEEE Global Conference on Signal and Information Processing, GlobalSIP 2014 - Atlanta, United States
Duration: Dec 3 2014Dec 5 2014

Publication series

Name2014 IEEE Global Conference on Signal and Information Processing, GlobalSIP 2014

Other

Other2014 IEEE Global Conference on Signal and Information Processing, GlobalSIP 2014
CountryUnited States
CityAtlanta
Period12/3/1412/5/14

All Science Journal Classification (ASJC) codes

  • Signal Processing
  • Information Systems

Keywords

  • Kernels
  • Large-scale learning
  • Low-rank constraints
  • Risk minimization
  • Tensor train format

Fingerprint Dive into the research topics of 'Learning multidimensional Fourier series with tensor trains'. Together they form a unique fingerprint.

  • Cite this

    Wahls, S., Koivunen, V., Poor, H. V., & Verhaegen, M. (2014). Learning multidimensional Fourier series with tensor trains. In 2014 IEEE Global Conference on Signal and Information Processing, GlobalSIP 2014 (pp. 394-398). [7032146] (2014 IEEE Global Conference on Signal and Information Processing, GlobalSIP 2014). Institute of Electrical and Electronics Engineers Inc.. https://doi.org/10.1109/GlobalSIP.2014.7032146