Rethinking CNN Architecture for Enhancing Decoding Performance of Motor Imagery-based EEG Signals

This repository is the official implementation of M-ShallowConvNet in pytorch-lightning style:

@article{kim2022rethinking,
  title={Rethinking CNN Architecture for Enhancing Decoding Performance of Motor Imagery-based EEG Signals},
  author={Kim, Sung-Jin and Lee, Dae-Hyeok and Lee, Seong-Whan},
  journal={IEEE Access},
  year={2022},
  publisher={IEEE}
}

Abstract

Brain–computer interface (BCI) is a technology that allows users to control computers by reflecting their intentions. Electroencephalogram (EEG)–based BCI has been developed because of its potential, however, its decoding performance is still insufficient to apply in the real–world environment. As deep learning methods achieve the significant performance in various domains, it has been applied in the EEG–based BCI domain. In particular, ShallowConvNet is one of the most widely used methods because of its robust decoding performance in multiple datasets. However, the model’s parameters have to be optimized to apply this model to various datasets each time, and we have also found some issues in architecture that disturb the stable training. In this paper, we highlight potential problems that might arise in ShallowConvNet and investigate the potential solutions. In addition, we propose a novel model, called M–ShallowConvNet, which solves the existing problems. The proposed model achieves the accuracies of 0.8164 and 0.8647 in datasets 2a and 2b of BCI Competition IV, respectively. Hence, we demonstrate that performance improvement can be achieved with only a few small modifications that resolve the problems of the conventional model.

1. Installation

1.1 Clone this repository

  $ git clone https://github.com/comojin1994/m-shallowconvnet.git
  $ cd m-shallowconvnet

1.2 Preparing data

After downloading the BCI Competition IV 2a & 2b data, revise the data's directory in the config files

#### Path ####
BASE_PATH: "REVISE HERE"
LOG_PATH: "./logs"
CKPT_PATH: "./checkpoints"

1.3 Environment setup

Create checkpoints and logs directory following the structure

    .
    ├── checkpoints/
    ├── logs/
    ├── ...
    ├── training.py
    ├── evaluation.py
    └── README.md

Build and access the docker container

  # The default CUDA version is 11.x
  # PLZ change the script if you use CUDA version of 10.x
  $ bash docker/start_docker.sh
  $ docker exec -it torch-server /bin/bash
  $ cd m-shallowconvnet

2. Performance

2.1 BCI Competition IV 2a

Subject No.	A1	A2	A3	A4	A5	A6	A7	A8	A9	Avg.
ShallowConvNet	0.833	0.517	0.931	0.743	0.747	0.625	0.816	0.847	0.823	0.765
M-ShallowConvNet	0.910	0.556	0.938	0.806	0.816	0.660	0.938	0.851	0.875	0.816

2.2 BCI Competition IV 2b

Subject No.	B1	B2	B3	B4	B5	B6	B7	B8	B9	Avg.
ShallowConvNet	0.759	0.689	0.762	0.963	0.997	0.841	0.925	0.916	0.844	0.855
M-ShallowConvNet	0.781	0.686	0.812	0.953	0.984	0.884	0.916	0.931	0.834	0.865

3. Feature visualization on BCI Competition IV 2a

t-SNE	Singular Value

4. Pre-trained weight

• BCI Competition IV 2a
• BCI Competition IV 2b

5. Training

Revise the bash script to fit the device environment

  # BCI Competition IV 2a
  $ bash script/bcicompet2a.sh

  # BCI Competition IV 2b
  $ bash script/bcicompet2b.sh

6. Evaluation

  $ python evaluation.py --config_name <CONFIG FILE> --ckpt_path <CKPT PATH>
  # CONFIG FILE: bcicompet2a_config, bcicompet2b_config

7. Device info

• GPU: Geforce RTX 3090 * 3
• CPU: Inter Core-X i9-10980XE