Convolutional Neural Network-based Transfer Learning and Knowledge Distillation using Multi-Subject Data in Motor Imagery BCI

In Brain Computer Interfaces (BCIs), with multiple recordings from different subjects in hand, a question arises regarding whether the knowledge of previously recorded subjects can be transferred to a new subject. In this study, we explore the possibility of transferring knowledge by using a convolutional network model trained on multiple subjects and fine-tuning the model on a small amount of data from a new subject, thus, reducing the calibration time by reducing the time needed to record data and train a model. Our results show a significant increase in 4-class classification accuracy on the BCI IV-2a competition data, even when a small subset of the data is provided for training.