Date of this Version
Deep learning is seldom used in the classification of electroencephalography (EEG) signals, despite achieving state of the art classification accuracies in other spatial and time series data. Instead, most research has continued to use manual feature extraction followed by a traditional classifier, such as SVMs or logistic regression. This is largely due to the low number of samples per experiment, high-dimensional nature of the data, and the difficulty in finding appropriate deep learning architectures for classification of EEG signals. In this thesis, several deep learning architectures are compared to traditional techniques for the classification of visually evoked EEG signals. We found that deep learning architectures using long short-term memory units (LSTMs) outperform traditional methods, while small convolutional architectures performed comparably to traditional methods. We also explored the use of transfer learning by training across multiple subjects and refining on a particular subject. This form of transfer learning further improved the classification accuracy of the deep learning models.
Advisors: Ashok Samal and Matthew Johnson