Mixed penalization for enhancing class separability of evoked related potentials in Brain-Computer Interfaces

RUBEN SPIES; PETERSON, VICTORIA

Mellieha

Conferencia; 9th International Conference "Inverse Problems: Modeling and Simulation"; 2018

The Eurasian Association on Inverse Problems

A braincomputer interface (BCI) is a system which provides an alternative way ofcommunication between the mind of a person and the outside world by using onlymeasured brain activity [1]. An efficient and non-invasive way of establish thecommunication is based on electroencephalography (EEG) and event-relatedpotentials (ERPs). An ERPs is an endogenous potential which results as aconsequence of an external and relevant stimuli [2]. Detecting the ERP signalimmersed in the ongoing EEG turns out to be an extremely hard and challenging binarypattern recognition problem. The Linear Discriminant Analysis (LDA) criterion is a well-known andwidely used dimensionality reduction tool in the context of supervisedclassification. Although the LDA generally results in good classificationperformances while keeping the solution simple, it fails when the number ofsamples is large relative to the number of observations in the given data. Thismeager classification performance is due to the poor estimation of covariancematrices used within LDA, which usually become ill-conditioned. Severalauthors, both from the BCI and the statistical research communities, haveproposed different regularized versions of LDA, showing always the advantagesof such tools. In thiswork we present a penalized version of the sparse discriminant analysis (SDA)[4], called generalized sparse discriminant analysis (GSDA) [5], for binaryclassification. This method inherits both the discriminative feature selectionand classification properties of SDA and it also improves SDA performancesthrough the addition of Kullback-Leibler class discrepancy information. TheGSDA method is designed to automatically select the optimal regularizationparameters by means of the L-hypersurface. Numerical experiments with two realERP-EEG datasets show that, on one hand, GSDA outperforms standard SDA in thesense of classification performance, sparsity and required computing time, and,on the other hand, it also yields better overall performances, as compared to mostof the state-of-the-art ERP classification algorithms, for single-trial ERPclassification, when insufficient training samples are available.