Subject Wise Motor Imagery Classification from EEG Data Using Transfer Learning

Machine learning (ML) has no doubt virtually helped in nearly all fields of life, including medical sciences. ML models are now being trained, tested and developed with the help of information gained from Electroencephalogram (EEG) Signals. Neural Networks (NN) are being used specifically in this regard to exploit their image classification ability. A special class of NN called Transfer Learning (TL), is used to enhance the capability of NNs. In this paper, EEG signals are extracted and used to classify Left or Right Motor Images of the brain using Inception V3 and VGG 16 models. We try to enhance the accuracy of these TL Models by exploiting a different methodology as compared to other available statistical methods available in the research community. For the said purpose, a dataset from Brain-Computer Interface (BCI) Competition IV 2b was used. EEG signals are extracted and transformed into Short Time Fourier Transform (STFT) images. These STFT images are labeled with either Left or Right Motor Imagery (MI) Class. The transfer learning models are trained using these STFT images and results are also compared with a state-of-the art research, implementing Capsule Networks.