Toward A Generalised Signal Processing Framework for Inter-Subject Associative BCI
Access status:
Embargoed
Type
ThesisThesis type
Doctor of PhilosophyAuthor/s
Saha, SimantoAbstract
An inter-subject associative BCI refers to a fully zero-training neural decoding algorithm when a group of users shares similar EEG features so that the algorithm can perform well for a user when trained on other users' data from the same group. However, only a few studies attempted ...
See moreAn inter-subject associative BCI refers to a fully zero-training neural decoding algorithm when a group of users shares similar EEG features so that the algorithm can perform well for a user when trained on other users' data from the same group. However, only a few studies attempted to quantify inter-subject associativity, i.e., identifying task-related EEG features that exhibit unchanged feature domains across subjects, minimizing the chance of covariate shift occurrence. Quantifying inter-subject associativity complements data-driven transfer learning and delivers good BCI performance without any training data. Data-driven conventional techniques, including common spatial pattern (CSP), are prone to overfitting and often demonstrate inconsistent classification accuracies. However, recently introduced convolutional neural network (CNN)-based architectures are deep learning techniques that inherently learn features from data with nonlinear activation functions while optimizing the models' predictive capabilities. A novel Bhattacharya distance-based predictor was developed for a CSP-based BCI classification framework. The CSP-based methods were compared with novel BCI classification pipelines utilizing a 1-dimensional convolutional neural network (1D-CNN) architecture. Various feature representation techniques, such as bandpass-filtered EEG signals, power spectral density (PSD) sequences, and bi-channel cross-power spectral density (CPSD) sequences, were used to train the proposed 1D-CNN architectures. The proposed methods were tested on motor imagery (MI) and speech classification tasks from EEG signals. Results implicated that 1D-CNN, utilizing time-domain EEG signals, produced better classification accuracies than frequency-embedded PSD or CPSD sequences and CSP-based methods for intra- and inter-subject MI classification. However, the proposed 1D-CNN with PSD sequences outperformed the results of time-domain EEG signals for intra-subject speech BCI.
See less
See moreAn inter-subject associative BCI refers to a fully zero-training neural decoding algorithm when a group of users shares similar EEG features so that the algorithm can perform well for a user when trained on other users' data from the same group. However, only a few studies attempted to quantify inter-subject associativity, i.e., identifying task-related EEG features that exhibit unchanged feature domains across subjects, minimizing the chance of covariate shift occurrence. Quantifying inter-subject associativity complements data-driven transfer learning and delivers good BCI performance without any training data. Data-driven conventional techniques, including common spatial pattern (CSP), are prone to overfitting and often demonstrate inconsistent classification accuracies. However, recently introduced convolutional neural network (CNN)-based architectures are deep learning techniques that inherently learn features from data with nonlinear activation functions while optimizing the models' predictive capabilities. A novel Bhattacharya distance-based predictor was developed for a CSP-based BCI classification framework. The CSP-based methods were compared with novel BCI classification pipelines utilizing a 1-dimensional convolutional neural network (1D-CNN) architecture. Various feature representation techniques, such as bandpass-filtered EEG signals, power spectral density (PSD) sequences, and bi-channel cross-power spectral density (CPSD) sequences, were used to train the proposed 1D-CNN architectures. The proposed methods were tested on motor imagery (MI) and speech classification tasks from EEG signals. Results implicated that 1D-CNN, utilizing time-domain EEG signals, produced better classification accuracies than frequency-embedded PSD or CPSD sequences and CSP-based methods for intra- and inter-subject MI classification. However, the proposed 1D-CNN with PSD sequences outperformed the results of time-domain EEG signals for intra-subject speech BCI.
See less
Date
2025Rights statement
The author retains copyright of this thesis. It may only be used for the purposes of research and study. It must not be used for any other purposes and may not be transmitted or shared with others without prior permission.Faculty/School
Faculty of Engineering, School of Biomedical EngineeringAwarding institution
The University of SydneyShare