Tikhonov Regularization Enhances EEG-based Spatial Filtering for Single Trial Regression

Abstract : In the field of Brain-Computer Interfaces (BCI), robust methods for the decoding of continuous brain states are of great interest as new application fields are arising. When capturing brain activity by an elec-troencephalogram (EEG), the Source Power Comodulation (SPoC) algorithm allows to compute spatial filters for the decoding of a continuous variable. However, dealing with high-dimensional EEG data that suffer from low signal-to-noise ratio, the method reveals instabilities for small training data sets and is prone to overfitting. In this paper, we introduce a framework for applying Tikhonov regularization to the SPoC approach in order to restrict the solution space of filters. Our findings show that an additional trace normalization of the included covariance matrices is a necessary prerequisite to tune the sensitivity of the resulting algorithm. In an offline analysis with data from N=18 subjects, the introduced trace normalized and Tihonov regularized SPoC variant (NTR-SPoC) outperforms the standard SPoC method for the majority of individuals. With this proof-of-concept study, a generalizable regularization framework for SPoC has been established which allows to implement a variety of different regularization strategies in the future.
Type de document :
Communication dans un congrès
2017 - 7th International Brain-Computer Interface Conference, Sep 2017, Graz, Austria. pp.1-6
Liste complète des métadonnées

Littérature citée [20 références]  Voir  Masquer  Télécharger

https://hal.inria.fr/hal-01655755
Contributeur : Fabien Lotte <>
Soumis le : lundi 11 décembre 2017 - 12:40:13
Dernière modification le : jeudi 11 janvier 2018 - 06:24:07

Fichier

submission_final_24042017.pdf
Fichiers produits par l'(les) auteur(s)

Identifiants

  • HAL Id : hal-01655755, version 2

Collections

Citation

Andreas Meinel, Fabien Lotte, Michael Tangermann. Tikhonov Regularization Enhances EEG-based Spatial Filtering for Single Trial Regression. 2017 - 7th International Brain-Computer Interface Conference, Sep 2017, Graz, Austria. pp.1-6. 〈hal-01655755v2〉

Partager

Métriques

Consultations de la notice

250

Téléchargements de fichiers

86