Computationally efficient statistical inference in Markovian models
Loading...
URL
Journal Title
Journal ISSN
Volume Title
School of Electrical Engineering |
Doctoral thesis (article-based)
| Defence date: 2024-08-26
Unless otherwise stated, all rights belong to the author. You may download, display and print this publication for Your own personal use. Commercial use is prohibited.
Authors
Date
Major/Subject
Mcode
Degree programme
Language
en
Pages
157 + app. 203
Series
Aalto University publication series DOCTORAL THESES, 135/2024
Abstract
Markovian systems are ubiquitous in nature, science, and engineering, to model the evolution of a system for which the future state of the system only depends on the past through the present state. These often appear as time series or stochastic processes, and when they are partially observed, they are known under the umbrella term of state-space models. Inferring the current state of the system from these partial, and often noisy, observations is a fundamental question in statistics and machine learning, and it is often solved using Bayesian inference methods that correct a prior belief on the state of the system through the likelihood of the observations. This perspective gives rise to typically recursive algorithms, which sequentially process the observations to slowly refine the estimate of the current state of the system. The most common of these algorithms are the Kalman filter and its extensions via linearisation procedures, and particle filtering methods, based on Monte Carlo. Another question, which often arises is that of the past state or past trajectory of the system, given all the observations. Furthermore, it may also be of interest to identify the model itself, whereby the most likely (or any other metric) model within a family is picked given the observations. In this thesis, we examine the three problems of Bayesian filtering, smoothing, and identification in the context of Markovian models, and we propose computationally efficient algorithms to solve them. In particular, we develop the parallelisation of the recursive structure of the filteringsmoothing algorithms, which, while optimal in a sequential setting, can be significantly sped up by using modern parallel computing architectures. This endeavour is tackled in both the context of particle approximations and Kalman-related methods. Another important aspect of the thesis is the use of gradient-based methods to perform inference in state-space models, taking several forms. One of these is the generalisation of the Metropolis-adjusted Langevin algorithm (MALA) and related algorithms to the context of particle and Kalman filters, and their implication for high-dimensional state inference. Another one is making particle filters differentiable by approximating the usual algorithm and then using the approximation to perform inference in statespace models using gradient-based methods. Finally, we also discuss the use of gradient-flows to perform automatic locally optimal filtering in state-space models. Some of these algorithms are de facto sequential and hardly parallelisable, but some instances can benefit from parallelisation, and we discuss the implications of this in terms of computational efficiency.Description
Supervising professor
Särkkä, Simo, Prof., Aalto University, Department of Electrical Engineering and Automation, FinlandThesis advisor
Särkkä, Simo, Prof., Aalto University, Department of Electrical Engineering and Automation, FinlandOther note
Parts
-
[Publication 1]: Adrien Corenflos, James Thornton, George Deligiannidis, Arnaud Doucet.Differentiable Particle Filtering via Entropy-Regularized Optimal Transport.In Proceedings of the 38th International Conference on MachineLearning, Volume 139, Pages 2100–2111, July 2021.
DOI: 10.48550/arXiv.2102.07850 View at publisher
-
[Publication 2]: Fatemeh Yaghoobi, Adrien Corenflos, Sakira Hassan, and Simo Sarkka.Parallel Iterated Extended and Sigma-Point Kalman Smoothers. In Proceedingsof the 2021 IEEE International Conference on Acoustics, Speechand Signal Processing (ICASSP), Pages 5350–5354, June 2021.
DOI: 10.1109/ICASSP39728.2021.9413364 View at publisher
-
[Publication 3]: Adrien Corenflos, Zheng Zhao, and Simo Sarkka. Temporal GaussianProcess Regression in Logarithmic Time. In Proceedings of the 2022International Conference on Information Fusion (FUSION), Linkoping,Sweden, Pages 1–5, July 2022.
DOI: 10.48550/arXiv.2102.09964 View at publisher
-
[Publication 4]: Adrien Corenflos, Nicolas Chopin, and Simo Sarkka. De-SequentializedMonte Carlo: a parallel-in-time particle smoother. Journal of MachineLearning Research, Volume 23, Number 283, Pages 1–39, August 2022.
DOI: 10.48550/arXiv.2202.02264 View at publisher
- [Publication 5]: Adrien Corenflos and Simo Sarkka. Auxiliary MCMC samplers for parallelisableinference in high-dimensional latent dynamical systems. Submittedto Electronic Journal of Statistics, July 2023.
-
[Publication 6]: Adrien Corenflos and Hany Abdulsamad. Variational Gaussian filteringvia Wasserstein gradient flows. In Proceedings of the 31st European SignalProcessing Conference (EUSIPCO), Helsinki, Finland, Pages 1838–1842,September 2023.
Full text in Acris/Aaltodoc: https://urn.fi/URN:NBN:fi:aalto-202401041218DOI: 10.23919/EUSIPCO58844.2023.10289853 View at publisher
-
[Publication 7]: Adrien Corenflos and Axel Finke. Particle-MALA and Particle-mGRAD:Gradient-based MCMC methods for high-dimensional state-space models.Submitted to Journal of Machine Learning Research, January 2024.
DOI: 10.48550/arXiv.2401.14868 View at publisher