Code for
| Methodology | Mean (SD) | Median | Range (Max-Min) |
|---|---|---|---|
| Subject-Independent CNN | 74.15 (±15.83) | 75.00 | 60.00 (100.00-40.00) |
| Subject-Independent Deep CNN | 84.19 (±9.98) | 84.50 | 47.50 (99.50-52.00) |
| Subject-Adaptive Deep CNN | 86.89 (±11.41) | 88.50 | 44.00 (100.00-56.00) |
| Subject-Independent Meta-Learning | 87.20 (±10.54) | 89.50 | 41.00 (100.00-59.00) |
| Subject | EEGNet | DeepConvNet | Transfer Learning | Meta-Learning | Transfer Meta-Learning |
|---|---|---|---|---|---|
| 1 | 30.00 | 23.33 | 32.00 | 24.67 | 30.00 |
| 2 | 30.41 | 26.32 | 29.47 | 24.74 | 32.11 |
| 3 | 30.56 | 26.15 | 32.31 | 26.92 | 35.38 |
| 4 | 23.75 | 26.84 | 27.89 | 23.16 | 30.00 |
| 5 | 30.00 | 25.79 | 30.00 | 25.79 | 30.00 |
| 6 | 27.78 | 22.89 | 32.53 | 21.08 | 32.53 |
| 7 | 30.41 | 24.74 | 28.42 | 27.89 | 28.95 |
| 8 | 32.11 | 26.67 | 28.00 | 26.00 | 29.33 |
| 9 | - | 23.16 | 27.37 | 28.95 | 31.58 |
| 10 | - | 26.32 | 34.21 | 26.32 | 31.58 |
| Average | 29.67 | 24.40 | 30.22 | 25.55 | 31.15 |
Motor Imagery Raw Dataset: Link
The pre-processing code of the dataset is provided by Kaishuo et al.: Link
Inner Speech Dataset: Link
Original Inner Speech Github: Link
Conditions = Motor Imagery
Classes = "Left", "Right"
Download the motor imagery raw dataset from the resources above, and save them to the same $source folder. To conserve space, you may only download files that ends with EEG_MI.mat.
The following command will read the raw dataset from the $source folder, and output the pre-processed data KU_mi_smt.h5 into the $target folder.
python preprocess_h5_smt.py $source $target
Run train_motor_LOSO.py
usage: python train_motor_LOSO.py [--META] [DATAPATH] [OUTPATH] [-gpu GPU] [-fold FOLD]
Training a subject-indepdendent meta-learning baseline model with cross validation for a single subject.
Positional Arguments:
DATAPATH Path for the pre-processed EEG signals file
OUTPATH Path to folder for saving the trained model and results in
Optional Arguments:
--meta Set to enable meta-learning, default meta-learning is switched off
-gpu GPU Set gpu to use, default is 0
-fold FOLD Set the fold number to determine subject for training a binary-class motor imagery classification model
To obtain baseline for all subjects, run train_motor_LOSO_all.py:
usage: python train_motor_LOSO_all.py [--META] [DATAPATH] [OUTPATH] [-gpu GPU]
Trains subject-indepdendent meta-learning baseline model with cross validation for all subjects from the KU dataset.
Positional Arguments:
DATAPATH Path for the pre-processed EEG signals file
OUTPATH Path to folder for saving the trained model and results in
Optional Arguments:
--meta Set to enable meta-learning, default meta-learning is switched off
-gpu GPU Set gpu to use, default is 0
Obtaining baseline models for subsequent adaptation, run eval_motor_base.py:
usage: python eval_motor_base.py [DATAPATH] [MODELPATH] [OUTPATH] [-gpu GPU]
Evaluate performance of models for each subject and selects models for subsequent transfer-learning adaptation.
Positional Arguments:
DATAPATH Path for the pre-processed EEG signals file
MODELPATH Path to folder containing the baseline models
OUTPATH Path to folder for saving the selected models and results in
Optional Arguments:
-gpu GPU Set gpu to use, default is 0
To perform adaptation on the selected baseline models, run train_motor_adapt_all.py:
usage: python train_motor_adapt_all.py [DATAPATH] [MODELPATH] [OUTPATH] [-scheme SCHEME] [-trfrate TRFRATE] [-lr LR] [-gpu GPU]
Evaluate performance of models for each subject and selects models for subsequent transfer-learning adaptation.
Positional Arguments:
DATAPATH Path for the pre-processed EEG signals file
MODELPATH Path to folder containing the selected baseline models for adaptation
OUTPATH Path to folder for saving the adaptation results in
Optional Arguments:
-scheme SCHEME Set scheme which determines layers of the model to be frozen
-trfrate TRFRATE Set amount of target subject data to be used for subject-adaptive transfer learning
-lr LR Set the learning rate of the transfer learning
-gpu GPU Set gpu to use, default is 0
Conditions = Inner Speech
Classes = "Arriba/Up", "Abajo/Down", "Derecha/Right", "Izquierda/Left"
Download the inner speech raw dataset from the resources above, save them to the save directory as the main folder.
To perform subject-independent meta-learning on chosen subject, run train_speech_LOSO.py. Meta learning may also be utilized.
usage: python train_speech_LOSO.py [ROOTDIR] [--META] [-subj SUBJ]
Trains a baseline classifier for a chosen subject either using baseline backpropagation or meta-learning.
Positional Arguments:
ROOTDIR Root directory path to the folder containing the EEG data
Optional Arguments:
--meta Set to enable meta-learning, default meta-learning is switched off
-subj SUBJ Set subject to perform transfer learning adaptation on
To train classifiers for all subjects, run train_speech_LOSO_all.py. Contains the same arguments as train_speech_LOSO.py except without the subject argument. Example usage to train meta-learning for all subjects:
python train_speech_LOSO_all.py ROOTDIR --meta
To perform subject-adaptive transfer learning on chosen subject, run train_speech_adapt.py. Meta learning may also be utilized.
usage: python train_adapt_all.py [ROOTDIR] [MODELPATH] [OUTPATH] [--META] [-scheme SCHEME] [-trfrate TRFRATE] [-lr LR] [-gpu GPU] [-subj SUBJ]
Performs subject-adaptive transfer learning on a subject with or without meta-learning.
Positional Arguments:
ROOTDIR Root directory path to the folder containing the EEG data
MODELPATH Path to folder containing the baseline models for adaptation
OUTPATH Path to folder for saving the adaptation results in
Optional Arguments:
--meta Set to enable meta-learning, default meta-learning is switched off
-scheme SCHEME Set scheme which determines layers of the model to be frozen
-trfrate TRFRATE Set amount of target subject data to be used for subject-adaptive transfer learning
-lr LR Set the learning rate of the transfer learning
-gpu GPU Set gpu to use, default is 0
-subj SUBJ Set subject to perform transfer learning adaptation on
To perform subject-adaptive transfer learning on all subjects, run train_speech_adapt_all.py. Meta learning may also be utilized. Contains the same arguments as train_speech_adapt.py except without the subject argument.
To recreate for meta-learning subject-adaptation, run:
python train_speech_adapt_all.py ROOTDIR MODELPATH OUTPATH --meta -scheme 1 -trfrate 40
For normal subject-adaptation:
python train_speech_adapt_all.py ROOTDIR MODELPATH OUTPATH -scheme 1 -trfrate 40
Please cite
@article{ng2024subject,
title={Subject-independent meta-learning framework towards optimal training of EEG-based classifiers},
author={Ng, Han Wei and Guan, Cuntai},
journal={Neural Networks},
pages={106108},
year={2024},
publisher={Elsevier}
}
Article Publication Link Subject-independent meta-learning framework towards optimal training of EEG-based classifiers
CBCR License 1.0
Copyright 2024 Centre for Brain Computing Research (CBCR)
Redistribution and use for non-commercial purpose in source and binary forms, with or without modification, are permitted provided that the following conditions are met:
-
Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer.
-
Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution.
-
Neither the name of the copyright holder nor the names of its contributors may be used to endorse or promote products derived from this software without specific prior written permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
- In the event that redistribution and/or use for commercial purpose in source or binary forms, with or without modification is required, please contact the contributor(s) of the work.