Decoding Brain Signals with Meta-learning

Abstract

Brain activities recorded while performing mental imagination of body motor parts are called motor imagery signals. In the field of Brain Computer Interface (BCI), it has been observed that motor imagery classification model trained for one person doesn’t fit well for others. And the reason for this being, Electroencephalogram (EEG) measurements recorded while performing motor imagery are different for every other person as everyone has slightly different foldings of cortex, functional map etc. To solve this problem, many researchers have proposed various conventional, and deep learning based classification models. To our knowledge, most of the works in this field train different models for different individuals. But it is not practical to train a model from scratch for every individual who will be using a real world BCI application. We propose a meta-learning based approach for motor imagery signal classification where a model is trained on a variety of learning tasks, such that it is capable of learning new tasks using only a small number of training samples. Thus only one model is required to be trained for all the subjects. We have conducted our experiments on the BCI competition IV-2b dataset consisting of 9 subjects performing left hand and right hand motor imagery task. The results signifies that subject specific calibration is a much better and optimal approach as compaired to subject specific training as the fine tuned meta learnt model outperforms subject specific trained models (Source code avaliable at https://github.com/RahulnKumar/EEG-Meta-Learning.).