Skip to Main content Skip to Navigation

Exploration of multivariate EEG /MEG signals using non-stationary models

Abstract : Independent Component Analysis (ICA) models a set of signals as linear combinations of independent sources. This analysis method plays a key role in electroencephalography (EEG) and magnetoencephalography (MEG) signal processing. Applied on such signals, it allows to isolate interesting brain sources, locate them, and separate them from artifacts. ICA belongs to the toolbox of many neuroscientists, and is a part of the processing pipeline of many research articles. Yet, the most widely used algorithms date back to the 90's. They are often quite slow, and stick to the standard ICA model, without more advanced features.The goal of this thesis is to develop practical ICA algorithms to help neuroscientists. We follow two axes. The first one is that of speed. We consider the optimization problems solved by two of the most widely used ICA algorithms by practitioners: Infomax and FastICA. We develop a novel technique based on preconditioning the L-BFGS algorithm with Hessian approximation. The resulting algorithm, Picard, is tailored for real data applications, where the independence assumption is never entirely true. On M/EEG data, it converges faster than the `historical' implementations.Another possibility to accelerate ICA is to use incremental methods, which process a few samples at a time instead of the whole dataset. Such methods have gained huge interest in the last years due to their ability to scale well to very large datasets. We propose an incremental algorithm for ICA, with important descent guarantees. As a consequence, the proposed algorithm is simple to use and does not have a critical and hard to tune parameter like a learning rate.In a second axis, we propose to incorporate noise in the ICA model. Such a model is notoriously hard to fit under the standard non-Gaussian hypothesis of ICA, and would render estimation extremely long. Instead, we rely on a spectral diversity assumption, which leads to a practical algorithm, SMICA. The noise model opens the door to new possibilities, like finer estimation of the sources, and use of ICA as a statistically sound dimension reduction technique. Thorough experiments on M/EEG datasets demonstrate the usefulness of this approach.All algorithms developed in this thesis are open-sourced and available online. The Picard algorithm is included in the largest M/EEG processing Python library, MNE and Matlab library, EEGlab.
Complete list of metadatas

Cited literature [160 references]  Display  Hide  Download
Contributor : Abes Star :  Contact
Submitted on : Friday, March 13, 2020 - 3:15:10 PM
Last modification on : Monday, October 12, 2020 - 7:41:51 PM
Long-term archiving on: : Sunday, June 14, 2020 - 2:14:50 PM


Version validated by the jury (STAR)


  • HAL Id : tel-02507788, version 1



Pierre Ablin. Exploration of multivariate EEG /MEG signals using non-stationary models. Machine Learning [stat.ML]. Université Paris Saclay (COmUE), 2019. English. ⟨NNT : 2019SACLT051⟩. ⟨tel-02507788⟩



Record views


Files downloads