Using Long Short- Term Memory Network for Recognizing Motor Imagery Tasks

Abstract

Classifying the electrocorticogram (ECoG) signals based on motor imagery (MI) is one of the important issues of the BCI systems. Deep learning approaches have been most popularly applied to learn representations and classify different types of data. However, the number of studies that modeling cognitive events from ECoG signals are very limited. In this paper, we propose a deep learning method to use long short-term memory (LSTM) recurrent neural networks for learning representations from ECoG and gradient boosting (GB) for classifying MI ECoG, and demonstrate its advantages. First, we transform multichannel ECoG time-series into the LSTM-GB model including sequential information. After that, we train an LSTM neural network to learn robust spatial-temporal representations. The subtle temporal dependencies of ECoG data streams can be extracted from LSTM with unique information processing mechanism. The LSTM features coupled with the GB classifier can yield the satisfactory accuracy of 100% on publicly available ECoG dataset. Experiments demonstrate that the proposed method can effectively recognize different MI tasks. Empirical evaluation on the MI classification tasks demonstrates significant improvements in classification accuracy over current state-of-the-art approaches in MI-based BCI field.