BiLSTM and SqueezeNet With Transfer Learning for EEG Motor Imagery Classification: Validation With Own Dataset

Abstract

Transfer Learning (TL) is a methodology that allows the re-train of a Machine Learning (ML) algorithm (like Neural Networks or NN’s) for a new task with the advantage of the previous training acquired knowledge; with this methodology, it is possible to train NNs for a new task even if the data is scarce. The present study uses this approach to train NNs to classify Electroencephalography (EEG) signals that include Movement/Imagery (MI), first with a publicly available data set and then using it to validate the training process of a small dataset of acquired data. The first part of the article describes the methodology for acquiring EEG signals that imitated the information found in the publicly available dataset Physionet Motor/Imagery. The second part compares the training process for NNs. The first NN is a Bidirectional Long-Short Term Memory (BiLSTM) trained from scratch with the Physionet dataset, and the second NN is a CNN called SqueezeNet trained following the TL method with the small acquired dataset, reaching an accuracy of 91.25% in the BiLSTM with the scratch method and an accuracy of 92.33% with the transfer learning method for the EEG MI signal classification.

Keywords

• BCI
• EEG signal processing
• machine learning
• transfer learning
• motor/imagery