Decomposable Common Spatial Patterns
Translated title
Decomposable Common Spatial Patterns: Applying Graphical Model Selection in the Brain-Computer Interface: Applying Graphical Model Selection in the Brain-Computer Interface
Authors
Andersen, Nikolaj ; Faarkrog, Jannik Frank ; Søgaard, Mikkel Holm
Term
4. term
Education
Publication year
2014
Submitted on
2014-06-10
Pages
12
Abstract
Hjerne-computer-grænseflader (BCI) bruger elektroencefalografi (EEG) til at tolke hjerneaktivitet. Men signalerne varierer både mellem personer og mellem optagelser, så systemerne skal ofte kalibreres med nye, mærkede data, hvilket er tidskrævende. Når antallet af EEG-kanaler vokser, kræves der også mere data (den såkaldte “dimensionalitetsforbandelse”). Vi adresserer dette ved at bruge modelselektion i urettede Gaussiske grafiske modeller til at reducere antallet af parametre, der skal estimeres. I denne ramme beskrives relationer mellem kanaler med en sparsom præcisionsmatrix, hvor nuller betyder betinget uafhængighed mellem kanaler. Med udgangspunkt i den udbredte metode Common Spatial Patterns (CSP) foreslår vi Decomposable CSP (DCSP), som afledes ud fra præcisionsmatricer i stedet for kovariansmatricer. Vi evaluerer tilgangen på både et eksisterende og et nyt datasæt og sammenligner med CSP. DCSP giver lavere fejlrater med relative reduktioner på op til 50 %. Vi finder også, at blot at bruge præcisionsbaserede features forbedrer ydeevnen, sandsynligvis fordi de er mere robuste over for outliers (ekstremværdier). Gevinsten ved færre parametre bliver større, jo flere kanaler der er.
Brain-computer interfaces (BCIs) use electroencephalography (EEG) to interpret brain activity. However, signals vary across people and recording sessions, so systems often need new labeled calibration data, which is time-consuming. As the number of EEG channels increases, the amount of data required also grows (the “curse of dimensionality”). We address this by using model selection in undirected Gaussian graphical models to reduce the number of parameters that must be estimated. In this framework, connections between channels are encoded in a sparse precision matrix, where zeros indicate conditional independence. Building on the standard Common Spatial Patterns (CSP) method, we propose Decomposable CSP (DCSP), derived from precision matrices instead of covariance matrices. We evaluate the approach on both an existing dataset and a new dataset and compare it to CSP. DCSP achieves lower error rates, with relative reductions of up to 50%. We also find that simply using precision-based features improves performance, likely because they are more robust to outliers. The benefits of having fewer parameters become larger as the number of channels increases.
[This abstract was generated with the help of AI]
Documents