RawConvNet: An End to End Network for MI-EEG Decoding with Attention Mechanism and No Preprocessing

In this paper, we propose RawConvNet for motor imagery electroencephalogram (MI-EEG) classification, an end-to-end deep learning model for MI-EEG decoding in BrainComputer Interface (BCI) with minimum pre-processing. The proposed network is built on top of LightConvNet architecture proposed in [1]. LightConvNet is made end-to-end by including temporal convolution, removing the filterbanks, and modifying temporal attention module. The modified model, RawConvNet, achieves a state-of-the-art average accuracy of 78.64% on the BCIC-IV-2a dataset in a 4-class motor imagery task. The code of our proposed model is made available at https://github.com/ExploringCodes/RawConvNet