Fig 1.
Simulation of swarm motion controlled by the proposed BCI system.
Figs (a) and (b) assign the task of imagining moving the left hand (class 1) and the right hand (class 3) to control the swarm density and the shape of the formation, respectively. Figs (c) and (d) assign the task of imagining saying the words “concentrate” (class 2) and “split” (class 4) to concentrate and split the swarm, respectively. Fig (e,f,g,h) show the system’s feedback to the corresponding imagery of the users.
Fig 2.
Training Procedure for a sub-model Relevance Vector Machine.
Fig 3.
Procedure of the proposed online learning mixture of RVM models.
Symbols ⊚ and ⊙ represent the In and Out connection respectively.
Fig 4.
Estimate weights wi for the mixture of RVM models.
Table 1.
Prediction accuracy at each run Di.
Table 2.
Quality of Confusion Matrix (QCM) at each run Di.
Fig 5.
Confusion matrix’s quality of subjects S1-S8 over runs D1 − D6.
Table 3.
P-value of Wilcoxon left-tail signed rank test on performance using Accuracy (Table 1) and QMC (Table 2).
Table 4.
Linear regression of QCM for each subject S1−8 over runs D1−4.
Table 5.
Separability score of data for each class pair.
(a)-(h) are the separability value of subject S1.
Table 6.
Separability score of data for each class pair.
(a)-(h) are the separability value of subject S2.
Table 7.
Separability score of data for each class pair.
(a)-(h) are the separability value of subject S3.
Table 8.
Separability score of data for each class pair.
(a)-(h) are the separability value of subject S4.
Table 9.
Separability score of data for each class pair.
(a)-(h) are the separability value of subject S5.
Table 10.
Separability score of data for each class pair.
(a)-(h) are the separability value of subject S6.
Table 11.
Separability score of data for each class pair.
(a)-(h) are the separability value of subject S7.
Table 12.
Separability score of data for each class pair.
(a)-(h) are the separability value of subject S8.
Table 13.
Range of separability score for each mental task pair taken from all subject’s runs.
Fig 6.
Multi-class Common Spatial Pattern topology.
The first 4 CSP patterns extracted from Run 1 (first row) and the run thats has the highest averaged separability (second row) of 8 subjects.
Fig 7.
t-SNE visualization of tangent vector feature for run 1 and run 6 of subject 5.
Fig 8.
Data instability for each class along the runs.
The legend notations of each class (C1-C4) are given in (a).