Abstract
Electroencephalography (EEG) is a non-invasive recording of the brain’s electrical activity. Due to its low cost, high temporal resolution, and noninvasive data collection, EEG has been widely used in various important applications and research areas. Among these, classifications of Motor Imagery (MI) EEG and Depression EEG are the two prominent research fields as they will significantly help improve the quality of life of millions of people worldwide. This thesis mainly focuses on improving the classification accuracy of MI-EEG and depression EEG using deep learning models.
To achieve this, an in-depth survey is conducted to evaluate and analyze the existing DL-based studies in MI-EEG classification. Apart from reproducing and objectively comparing the performance of current state-of-the-art deep learning models in MI-EEG classification, the survey also investigates the effectiveness of some key design factors in representative network architectures and identifies open issues in the field. A main open issue in MI-EEG classification is data distribution discrepancy between the training and test data, which often leads to poor model generalization. To address this issue, this thesis develops a novel parallel feature fusion multi-task learning model where MI-EEG classification and a related task (input reconstruction) are trained jointly to learn generalized feature representations. Through gate control modules, two separate task branches in the proposed model interact with each other, and ensure useful information is selectively transmitted between the two tasks. Experimental results demonstrate that the proposed model outperforms state-of-the-art models on two public MI-EEG datasets under most metrics, especially when the distribution discrepancy between training and test data is large, e.g., in cross-subject evaluation.
This thesis also focuses on the classification of depression EEG. A main issue in this task is data scarcity. To address this issue, this thesis proposes a novel heterogeneous few-shot learning framework where multiple heterogeneous depression EEG datasets can be effectively learned simultaneously. The proposed model consists of an Inference Network and a Prediction Network. The Inference Network learns generalized EEG features from support data and encodes corresponding label information. The Prediction Network classifies query data by utilizing the EEG features and label information learned from the Inference Network. Experimental results demonstrate the superior performance of the proposed model in depression EEG classification under several experimental conditions, especially when the amount of training data from the target dataset is limited. A channel occlusion-based interpretability analysis approach has also been developed and adopted to identify important brain regions from the proposed model. According to our experiments, the brain regions this approach identifies are consistent with some existing neuroscience research findings.
In conclusion, this thesis verifies effective multi-task learning can alleviate data distribution discrepancy by jointly learning on multiple related tasks to improve generalization, and sophisticated heterogeneous few-shot learning can alleviate data scarcity by effectively utilizing other heterogeneous datasets for training. Both of them illustrate their potential to break the current limits on EEG classification. In future, I will further explore which auxiliary tasks are the best for MTL and how to develop more effective models to learn heterogeneous data in EEG classification.