Dagois, Elise
(2019)
Transfer learning for a multimodal hybrid EEG-FTCD Brain-Computer Interface.
Master's Thesis, University of Pittsburgh.
(Unpublished)
Abstract
Transfer learning has been used to overcome the limitations of machine learning in Brain-Computer Interface (BCI) applications. Transfer learning aims to provide higher performance than no-transfer machine learning when only a limited number of training data is available and can consequently reduce training and calibration requirements. BCI systems are designed to provide communication and control tools for individuals with limited speech and physical abilities (LSPA). Most noninvasive BCI systems are based on Electroencephalogram (EEG) because of EEG \textquotesingle s cost effectiveness and portability. However, EEG signals present low signal-to-noise ratio and nonstationarity due to background brain activity. Such a behavior may decrease the global performance of the system. To overcome the disadvantages of EEG signals, in our previous work, we developed two different multi-modal BCI systems based on EEG and functional transcranial Doppler (fTCD), a cerebral flood velocity measure. These two multi-modal systems that combine EEG and fTCD signals aim to reduce performance degradation obtained when EEG was the only BCI modality. One of the systems is based on steady state evoked potentials and the other one is designed using motor imagery paradigms. Our results have shown that such a hybrid system outperforms EEG only BCIs. However, both systems require significant amount of training data for personalized design which could be tiresome for the target population. In this study, we extend these systems by performing a new transfer learning algorithm and we demonstrate the corresponding algorithm on the three different binary classification tasks for both BCIs in order to reduce the calibration requirements. Performing experiments with healthy participants, we collected EEG and fTCD data using both BCI systems. In order to apply transfer learning and to reduce the calibration requirements for BCIs, for each participant, we identify the most informative datasets from the rest of the participants based on probabilistic similarities between the class conditional distributions and increase the training set from this data. We demonstrate that transfer learning reduces the calibration requirements up to \%87.5 for BCI systems. Also, through comparison between different classifiers LDA, QDA, and SVM, we observe that QDA achieves the higher difference between transfer learning and no transfer accuracy.
Share
| Citation/Export: |
|
| Social Networking: |
|
Details
| Item Type: |
University of Pittsburgh ETD
|
| Status: |
Unpublished |
| Creators/Authors: |
|
| ETD Committee: |
|
| Date: |
23 January 2019 |
| Date Type: |
Publication |
| Defense Date: |
26 November 2018 |
| Approval Date: |
23 January 2019 |
| Submission Date: |
26 November 2018 |
| Access Restriction: |
No restriction; Release the ETD for access worldwide immediately. |
| Number of Pages: |
47 |
| Institution: |
University of Pittsburgh |
| Schools and Programs: |
Swanson School of Engineering > Electrical and Computer Engineering |
| Degree: |
MS - Master of Science |
| Thesis Type: |
Master's Thesis |
| Refereed: |
Yes |
| Uncontrolled Keywords: |
Transfer Learning, Hybrid Brain Computer Interfaces, Electroencephalogram, Functional Transcranial Doppler Ultrasound, Distance Measures, Machine Learning. |
| Date Deposited: |
23 Jan 2019 15:49 |
| Last Modified: |
23 Jan 2019 15:49 |
| URI: |
http://d-scholarship.pitt.edu/id/eprint/35566 |
Metrics
Monthly Views for the past 3 years
Plum Analytics
Actions (login required)
 |
View Item |
![[feed]](/images/feed-icon-32x32.png){kind=icon}