English

Gaussian process deconvolution

Machine Learning 2023-07-19 v2 Machine Learning Signal Processing

Abstract

Let us consider the deconvolution problem, that is, to recover a latent source x()x(\cdot) from the observations y=[y1,,yN]\mathbf{y} = [y_1,\ldots,y_N] of a convolution process y=xh+ηy = x\star h + \eta, where η\eta is an additive noise, the observations in y\mathbf{y} might have missing parts with respect to yy, and the filter hh could be unknown. We propose a novel strategy to address this task when xx is a continuous-time signal: we adopt a Gaussian process (GP) prior on the source xx, which allows for closed-form Bayesian nonparametric deconvolution. We first analyse the direct model to establish the conditions under which the model is well defined. Then, we turn to the inverse problem, where we study i) some necessary conditions under which Bayesian deconvolution is feasible, and ii) to which extent the filter hh can be learnt from data or approximated for the blind deconvolution case. The proposed approach, termed Gaussian process deconvolution (GPDC) is compared to other deconvolution methods conceptually, via illustrative examples, and using real-world datasets.

Keywords

Cite

@article{arxiv.2305.04871,
  title  = {Gaussian process deconvolution},
  author = {Felipe Tobar and Arnaud Robert and Jorge F. Silva},
  journal= {arXiv preprint arXiv:2305.04871},
  year   = {2023}
}

Comments

Accepted at Proceedings of the Royal Society A