This is the official repository to the paper "EDPNet: An Efficient Dual Prototype Network for Motor Imagery EEG".
- Inspired by clinical prior knowledge of EEG-MI and human brain recognition mechanisms, we propose a high performance, lightweight, and interpretable MI-EEG decoding model EDPNet. The EDPNet simultaneously considers and overcomes three major challenges in MI-BCIs.
- To extract highly discriminative features from EEG signals, we design two novel modules, ASSF and MVP, for the feature extractor of EDPNet. The ASSF module extracts effective spatial-spectral features, and the MVP module extracts powerful multi-scale temporal features.
- To overcome the small-sample issue of MI tasks, we propose a novel DPL approach to optimize the distribution of features and prototypes, aiming to obtain a robust feature space. This enhances the generalization capability and classification performance of our EDPNet.
- We conduct experiments on three benchmark public datasets to evaluate the superiority of the proposed EDPNet against state-of-the-art (SOTA) MI decoding methods Additionally, comprehensive ablation experiments and visual analysis demonstrate the effectiveness and interpretability of each module in the proposed EDPNet.
- Python 3.10
- Pytorch 2.12
In the following datasets we have used the official criteria for dividing the training and test sets:
- BCI_competition_IV 2a -acc 84.11%
- BCI_competition_IV 2b -acc 86.65%
- BCI_competition_III IVa -acc 82.03%
If you have any questions, please feel free to email [email protected].