Hilbert Space Methods in Infinite-Dimensional Kalman Filtering

Arno SolinAaltodoc 2012

Abstract

Many physical and biological processes include both spatial and temporal features. Spatio-temporal modeling under the machine learning paradigm of Gaussian process (GP) regression has demonstrated prominent results. However, the appealing Bayesian treatment by GP regression is often difficult in practical problems due to computational complexity.\nIn this thesis, methods for writing spatio-temporal Gaussian process regression as infinite-dimensional Kalman filtering and Rauch - Tung - Striebel smoothing problems are presented. These scale linearly with respect to the number of time steps as opposed to the cubic scaling of the direct GP solution. Spatio-temporal covariance functions are formulated as infinite-dimensional stochastic differential equations. Furthermore, it is presented how infinite-dimensional models can be combined with a finite number of observations to an approximative solution. For this, a truncated eigenfunction expansion of the Laplace operator is formed in various domains, of which the n-dimensional hypercube and hypersphere are explicitly written out.\nThe approach in this thesis is primarily application-driven, and therefore three real-world case studies are presented as proof of concept. The feasibility of infinite-dimensional Kalman filtering is demonstrated by forming a spatio-temporal resonator model which is applied to temperature data in two spatial dimensions, and a novel way of modeling the space{time structure of physiological noise in functional brain imaging data is considered in both two and three spatial dimensions.
BibTeX
@inproceedings{solin2012hilbert,
  author = {Arno Solin},
  title = {{Hilbert Space Methods in Infinite-Dimensional Kalman Filtering}},
  booktitle = {Aaltodoc},
  year = {2012},
}

Loading related content...

Figure