Defence of doctoral thesis in the field of Bioelectronics and Instrumentation, M.Sc. Zheng Zhao
M.Sc. Zheng Zhao will defend the thesis "State-space deep Gaussian processes with applications" on 10 December 2021 at 12:00 in Aalto University School of Electrical Engineering, Department of Electrical Engineering and Automation, in lecture hall AS1, Maarintie 8, Espoo, and online in Zoom.
Opponent: Prof. Manfred Opper, University of Birmingham, UK
Custos: Prof. Simo Särkkä, Aalto University School of Electrical Engineering, Department of Electrical Engineering and Automation
The public defense will be organized via remote technology. Follow defence: https://aalto.zoom.us/j/67529212279
Zoom Quick Guide: https://www.aalto.fi/en/services/zoom-quick-guide
In the defence arranged at campus, the organiser may check the COVID19 certificate of the participants, depending on the number of participants present (more details at: https://www.aalto.fi/en/study-at-aalto/being-a-doctoral-student-at-aalto --> Public defences).
Thesis available for public display at: https://aaltodoc.aalto.fi/doc_public/eonly/riiputus/
Doctoral theses in the School of Electrical Engineering: https://aaltodoc.aalto.fi/handle/123456789/53
Learning unknown latent functions from data (i.e., regression) is a common problem in control, signal processing, statistics, and machine learning, for example, estimating the positions of targets from their radar measurements, or estimating stock price from company sales. This thesis presents a class of general-purpose probabilistic models -- state-space deep Gaussian processes (SS-DGPs) -- that are useful for modelling such unknown functions (as prior) in a number of applications. In particular, these models are capable of tackling irregular latent functions that have time-varying features (e.g., gravitational waves have time-varying frequency; Doppler effect can also change the frequency of signal in time). Moreover, thanks to the Markov property of these models, their regression problem can be solved efficiently in linear computational time with respect to the number of data points.
This thesis also presents a statistical Taylor expansion-based method for continuous-discrete Gaussian filtering and smoothing problems. This method can give closed-form approximations to the statistical quantities (e.g., mean and covariance) of models' stochastic dynamics in arbitary precision. This method is also useful for discretising stochastic differential equations and SS-DGPs.
Finally, this thesis features a few applications of state-space (deep) Gaussian processes. This includes, for example, system identification and spectro-temporal signal analysis.
Contact information of doctoral candidate: