Subject-independent ERP-based brain-computer interface using adaptive and ensemble learning

Publication Type:
Issue Date:
Full metadata record
The breakthrough of deep learning in recent years opens up a wide range of applications, mostly from computer vision to brain-computer interface (BCI), which is the research topic of this thesis. In our study, we propose a unified framework of the subject-independent event-related potential (ERP) based BCI. In other words, we attempt to overcome a major challenge of discrepancies in ERP patterns across different subjects by employing a deep learning technique, accompanied with different strategies of bagging-stacking ensemble, dynamic stopping, and adaptive incremental learning. The main contributions of this thesis are summarised as follows: (1) Employ the subject-adversarial neural network (SANN) to learn the optimal representation for the original preprocessed ERP features. This network serves as a feature encoder which attempts to find a better feature space that is subject-independent for the testing procedure. (2) Employ three machine learning algorithm as base learners: support vector machine (SVM), Fisher’s discriminant analysis (FDA), and fully-connected feedforward neural network (NN). The soft scores of these base learners serves as inputs for the ensemble strategy which is comprised of two widely-used techniques: bagging and stacking. The goal of the whole ensemble strategy is to perform the binary classification problem of ERP trials with lower generalisation error and higher accuracy. (3) Employ two post-processing tasks for the P300-Speller (P3S) to enhance the information transfer rate, namely dynamic stopping (DS), and adaptive (or incremental) learning (AL). DS is performed to let the system produce the subject’s output whenever the algorithm is sufficiently confident about its decision, while AL is used to reinforce the existing classifier by analytically integrating newly-classified samples into the classifier’s decision function in real-time. The research of the thesis is motivated by the strong representative characteristics of ERP features in EEG signal. The distinctive properties of ERP, especially P300 component in this thesis, and its variability across multiple subjects, can be well-exploited bythe high complexity and employment of deep adversarial neural network framework. The robust representation features output of this network are served as learning features for the subsequent step of ensemble learning by multiple machine learning algorithms. Finally, some post-processing and enhancement methods specifically proprietary to this thesis such as dynamic stopping and incremental learning are applied to yield an additional boost in our performance.
Please use this identifier to cite or link to this item: