SST-DPN

This is the official repository to the paper "A Spatial-Spectral and Temporal Dual Prototype Network for Motor Imagery Brain-Computer Interface".

Abstract

• To extract powerful spatial-spectral features, we design a lightweight attention mechanism that explicitly models the relationships among multiple channels in the spatial-spectral dimension. This method enables finer-grained spatial feature modeling, highlighting key spatial-spectral channels for the current MI task.
• To capture long-term temporal features from high temporal resolution EEG signals, we develop a Multi-scale Variance Pooling (MVP) module with large kernels. Compared to commonly used transformers, the MVP module is parameter-free and computationally efficient. Extensive experiments show that MVP outperforms transformers, indicating its potential as an alternative for capturing long-term temporal features in EEG signal decoding and real-time BCI applications.
• To overcome the small-sample issue, we propose a novel Dual Prototype Learning (DPL) method to optimize feature space distribution, making same-class features more compact and different-class features more separated. The DPL acts as a regularization technique, enhancing the model’s generalization ability and classification performance. Furthermore, the DPL can be easily integrated with existing advanced methods, serving as a general approach to enhance model performance. To the best of our knowledge, this paper is the first to apply the prototype learning to EEG-MI decoding. We believe it offers valuable insights that will inspire further advancements in the field.
• We conduct experiments on three benchmark public datasets to evaluate the superiority of the proposed SST-DPN 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 method.

Requirements:

• python 3.10
• pytorch 2.12
• braindecode 0.8.1
• moabb 1.1.0

Data download and preprocessing

All data will be downloaded automatically except for the BCI3-4A dataset. Download the BCI3-4A dataset and put all files in the directory defined in load_data.py.

Rusults and Visualization

In the following datasets we have used the official criteria for dividing the training and test sets:

• BCI4-2A -acc 84.11%
• BCI4-2B -acc 86.65%
• BCI3-4A -acc 82.03%

Acknowledgments

We are deeply grateful to Martin for providing clear and easily executable code in the channel-attention repository. In our paper, we referenced the code and results from channel-attention to ensure the reliability of our reproductions of the baseline methods.

We also appreciate the braindecode library for providing convenient tools for data downloading and preprocessing.

Contact

If you have any questions, please feel free to email hancan@sjtu.edu.cn.