Oren Shriki

EEG, Electroencephalography, is the acquisition and decoding of electric brain signals. The data acquired from EEG scans can be put to use in many fields, including seizure prediction, treatment of mental illness, brain-computer interfaces (BCIs) and more. Recent advances in deep learning (DL) in fields of image classification and natural language processing have motivated researchers to apply DL for classification of EEG signals as well. One major caveat in DL is the amount of human effort and expertise required for the development of efficient and effective neural network architectures. Neural architecture search algorithms are used to automatically find good enough neural network architectures for a problem and dataset at hand. In this research, we employ genetic algorithms for optimizing neural network architectures for multiple tasks related to EEG processing while addressing two unique challenges related to EEG: (1) small amounts of labeled EEG data per subject, and (2) high diversity of EEG signal patterns across subjects. Neural network architectures produced during this study successfully compete with state of the art architectures published in the literature. Particularly successful are architectures optimized for all (human) subjects, with evolution and training performed on a mixed dataset including all subjects’ data.