Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Source Data Selection for Brain-Computer Interfaces based on Simple Features

Frida Heskebeck, Carolina Bergeling, Bo Bernhardsson

Abstract

This paper demonstrates that simple features available during the calibration of a brain-computer interface can be utilized for source data selection to improve the performance of the brain-computer interface for a new target user through transfer learning. To support this, a public motor imagery dataset is used for analysis, and a method called the Transfer Performance Predictor method is presented. The simple features are based on the covariance matrices of the data and the Riemannian distance between them. The Transfer Performance Predictor method outperforms other source data selection methods as it selects source data that gives a better transfer learning performance for the target users.

Source Data Selection for Brain-Computer Interfaces based on Simple Features

Abstract

This paper demonstrates that simple features available during the calibration of a brain-computer interface can be utilized for source data selection to improve the performance of the brain-computer interface for a new target user through transfer learning. To support this, a public motor imagery dataset is used for analysis, and a method called the Transfer Performance Predictor method is presented. The simple features are based on the covariance matrices of the data and the Riemannian distance between them. The Transfer Performance Predictor method outperforms other source data selection methods as it selects source data that gives a better transfer learning performance for the target users.
Paper Structure (20 sections, 7 equations, 3 figures, 1 table)

This paper contains 20 sections, 7 equations, 3 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related work
  3. Background
  4. EEG-based motor imagery BCIs
  5. Classification of BCI data using Riemannian Geometry
  6. Transfer learning with the RPA method
  7. Material and Method
  8. Dataset and Preprocessing
  9. MI classification with transfer learning
  10. The Transfer Performance Predictor (TPP) method
  11. Source data selection
  12. Results and Discussion
  13. MI classification with transfer learning
  14. Source data selection
  15. Conclusions and Future Research
  16. ...and 5 more sections

Figures (3)

  • Figure 1: Accuracy of MI classification after using the RPA transfer learning method, average over 5-folds. The rows are target users, and the columns are source users. The colors correspond to the cross-subject MI classification accuracy for each combination of target and source users. The highlighted cells are the intra-subject accuracies. For an arbitrary row, we can see that the transfer learning accuracy differs a lot depending on which source data is used, which is the reason why source data selection for BCIs is important. The rows are sorted in decreasing order of the row sum, and columns are sorted in decreasing order of the column sum. An interpretation of this sorting is that target users that benefit from many source users are found at the top (good students), and source users that are beneficial for many target users are found to the left (good teachers).
  • Figure 2: Comparision of different source data selection methods. The colors and numbers represent the mean difference in transfer learning accuracy between the method in the row and the method in the column. The matrix is thus skew-symmetric. A higher number means that the method in the row, on average, selects source data that gives each target user a higher transfer learning accuracy than the method in the column. The highlighted cells indicate methods where the difference between the transfer learning accuracies for each target user is not statistically different from 0, meaning that one cannot say that one method is better. As expected, the Oracle method is the best.
  • Figure 3: Comparision of the source data selection methods for different numbers of source data candidates. The y-axis shows the method's performance compared to the Oracle performance. The x-axis shows the number of candidates suggested by each method for source selection. The figure shows the same numbers for six source candidates as the rightmost column in \ref{['fig:baseline_comparision']}. The more candidates each method can suggest, the closer to Oracle performance. The TPP method outperforms the other for three or more suggested candidates.