EMPT: a sparsity Transformer for EEG-based motor imagery recognition

This study introduces a new deep learning model, EMPT, for decoding EEG data related to motor imagery in patients with spinal cord injury. EMPT combines a Transformer neural network with a Mixture of Experts (MoE) layer and a ProbSparse Self-attention mechanism. The model aims to improve the accuracy of motor imagery recognition by introducing sparsity to the Transformer network, making it more applicable to EEG datasets. The MoE layer and ProbSparse Self-attention help the model to focus on the most relevant features in the EEG data, enhancing its performance. EMPT achieves an accuracy of 95.24% on the MI EEG dataset for patients with spinal cord injury, outperforming other state-of-the-art methods. This suggests that EMPT is a promising approach for decoding EEG data and enabling human-computer interaction for individuals with motor impairments.

• 1
  EMPT achieves an accuracy of 95.24% on the MI EEG dataset for patients with spinal cord injury.
• 2
  The MoE layer and ProbSparse Self-attention enhance the applicability of the Transformer network on EEG datasets by introducing sparsity.
• 3
  Visualisation experiments show that the MoE layer can effectively perform dynamic sub-model selection for individual subjects, achieving model sparsity.