A Deep Transfer Learning Training Strategy For Inter-Subject Classification of EEG Signal

Electroencephalogram(EEG) signals are generally available in small limited quantities, and there are considerable variabilities between individual and recording sessions. Thus it is crucial to obtain a model capable of finding discriminative features invariant to these differences. For this purpose, Convolutional Neural Networks (CNN) proved to be suited for detecting Spatio-temporal features that are discriminative for these signals. This paper introduces Stage-training as a variation of fine-tuning to train end-to-end CNN models explicitly designed for learning discriminative features from EEG recordings. The Stage-training approach specifically addresses the challenges of finding a model invariant to inter-subject differences when fine-tuning fails due to the particular structure of these models. Using this approach, which conducts multiple stages of fine-tuning in transfer learning, we significantly increased the overall accuracy of the EEGNet by 9%. Our strategy's implementation can be found at: https://github.com/mj-sam/stage-trans.