Use this URL to cite or link to this record in EThOS: https://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.789555
Title: Novel transfer learning approaches for improving brain computer interfaces
Author: Azab, Ahmed
ISNI:       0000 0004 8501 462X
Awarding Body: University of Sheffield
Current Institution: University of Sheffield
Date of Award: 2019
Availability of Full Text:
Access from EThOS:
Access from Institution:
Abstract:
Despite several recent advances, most of the electroencephalogram(EEG)-based brain-computer interface (BCI) applications are still limited to the laboratory due to their long calibration time. Due to considerable inter-subject/inter-session and intra-session variations, a time-consuming and fatiguing calibration phase is typically conducted at the beginning of each new session to acquire sufficient labelled training data to train the subject-specific BCI model. This thesis focuses on developing reliable machine learning algorithms and approaches that reduce BCI calibration time while keeping accuracy in an acceptable range. Calibration time could be reduced via transfer learning approaches where data from other sessions or subjects are mined and used to compensate for the lack of labelled data from the current user or session. In BCI, transfer learning can be applied on either raw EEG, feature or classification domains. In this thesis, firstly, a novel weighted transfer learning approach is proposed in the classification domain to improve the MI-based BCI performance when only few subject-specific trials are available for training. Transfer learning techniques should be applied in a different domain before the classification domain to improve the classification accuracy for subjects whom their subject-specific features for different classes are not separable. Thus, secondly, this thesis proposes a novel regularized common spatial patterns framework based on dynamic time warping and transfer learning (DTW-R-CSP) in raw EEG and feature domains. In previous transfer learning approaches, it is hypothesised that there are enough labelled trials available from the previous subjects or sessions. However, in the case when there are no labelled trials available from other subjects or sessions, domain adaptation transfer learning could potentially mitigate problems of having small training size by reducing variations between the testing and training trials. Thus, to deal with non-stationarity between training and testing trials, a novel ensemble adaptation framework with temporal alignment is proposed.
Supervisor: Arvaneh, Mahanz ; Mihaylova, Lyudmila Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID: uk.bl.ethos.789555  DOI: Not available
Share: