anp-scp / fast_bci

Pytorch code for the paper Evaluating Fast Adaptability of Neural Networksfor Brain-Computer Interface [IJCNN 24].

Fast BCI

Pytorch code for the paper Evaluating Fast Adaptability of Neural Networksfor Brain-Computer Interface [Accepted in IJCNN 24].

Open docs/index.html in a browser for web based documentation.

Abstract

Electroencephalography (EEG) classification is a versatile and portable technique for building non-invasive Brain-computer Interfaces (BCI). However, the classifiers that decode cognitive states from EEG brain data perform poorly when tested on newer domains, such as tasks or individuals absent during model training. Researchers have recently used complex strategies like Model-agnostic meta-learning (MAML) for domain adaptation. Nevertheless, there is a need for an evaluation strategy to evaluate the fast adaptability of the models, as this characteristic is essential for real-life BCI applications for quick calibration. We used motor movement and imaginary signals as input to Convolutional Neural Networks (CNN) based classifier for the experiments. Datasets with EEG signals typically have fewer examples and higher time resolution. Even though batch-normalization is preferred for Convolutional Neural Networks (CNN), we empirically show that layer-normalization can improve the adaptability of CNN-based EEG classifiers with not more than ten fine-tuning steps. In summary, the present work (i) proposes a simple strategy to evaluate fast adaptability, and (ii) empirically demonstrate fast adaptability across individuals as well as across tasks with simple transfer learning as compared to MAML approach.

Setup

Pre-requisite

1. Install conda
2. Install dependencies:

conda env create -f environment.yml

3. Activate the conda environment

conda activate fast_bci

4. Go to root directory of the code
5. Install the package in development mode

conda develop .

6. Download the data by running the following command. Running this for the first time may ask for a path for MNE_DATA. Set the desired path and continue.

python3 download_data.py

Directory structure

Terminology Alert

1. Subject: In BCI wrold, a person for whom EEG is recorded is refered as a subject.
2. In BCI world, a task can be an activity or human body movement. For example: moving hands
3. baseline: In the code, baseline refers to tranfer learning. In the beginning of the experiment, we thought that MAML will work better and named transfer learning to baseline. But, later we found that transfer learning works better.

Directory	Description
metalearning	Module for MAML related APIs
baseline	Module for transfer learning related APIs
across_subject	Experiments for across individual adaptability
across_task	Experiments for across activity adaptability

NOTE:

• Web based API docs for metalearning module is available in docs/metalearning/index.html.
• Web based API docs for baseline module is available in docs/baseline/index.html.

Scripts

1. Training CNN model using MAML: across_subject/train.py
2. Testing CNN model trained using MAML on new individuals: across_subject/test.py
3. Training CNN model using transfer learning: across_subject/baseline_train.py
4. Testing CNN model trained using transfer learning on new individuals: across_subject/baseline_test.py
5. Testing CNN model trained using MAML on new activities: across_task/test.py
6. Testing CNN model trained using transfer learning on new activities: across_task/baseline_test.py

The parameters of the scripts are defined in params.yaml in the respective directories.

Label mapping

We map EEG data with labels using a parameter called label_mapping. The annotation of EEG data is provided in the homepage of Physionet's EEG Motor Movement/Imagery Dataset.

Here, we describe how to map labels with data in the dataset.

For example, we need to perform binary classification for open and close left vs right fist (Task 1). Then, as per the annotation, we would need data with code T1 at runs 3,7,11 for left fist and data with code T2 at runs 3,7,11 for right fist. The following YAML snippet labels data for left fist as label 0 and data for right fist as label 1:

  label_mapping:
  # Mapping of labels and task/activity in Physionet's dataset
    0:
    - - 3
      - 7
      - 11
    - - T1
    1:
    - - 3
      - 7
      - 11
    - - T2

Experiments

Batch norm vs layer norm for adaptability

Use following values of parameters in across_subject/params.yaml under block across_subject_baseline for hyperparamter tuning:

1. lr --> [0.01, 0.001]
2. batch_size --> [16, 32, 64]
3. norm --> [layer, batch]

For training run following command inside across_subject dir:

python3 baseline_train.py

For testing run following command inside across_subject dir:

python3 baseline_test.py

NOTE:

Running the script once, will create a model for one hyperparameter set.

Aross individual adaptability

For MAML training, use following values of parameters in across_subject/params.yaml under block across_subject for hyperparamter tuning:

1. adapt_lr --> [0.01, 0.001]
2. meta_lr --> [0.01, 0.001]
3. adapt_steps --> [5, 10]

For training run following command inside across_subject dir:

python3 train.py

For testing run following command inside across_subject dir:

python3 test.py

For transfer learning, use following values of parameters in across_subject/params.yaml under block across_subject_baseline for hyperparamter tuning:

1. lr = [0.01, 0.001]
2. batch_size = [16, 32, 64]

For training run following command inside across_subject dir:

python3 baseline_train.py

For testing run following command inside across_subject dir:

python3 baseline_test.py

Across task adaptability

To evaluate the adaptability of the model on newer activities, we use the models traind during Aross individual adaptability. We copy the model trained for one activity saved at across_subject/models dir to across_task/models dir. We use the test scripts similar to ones in across_subject for testing.

To test on new activity, change the label_mapping in across_task/params.yaml for a new activity and run the respective scripts in across_task dir.

For evaluation of MAML, use across_task/test.py and for transfer learning use across_task/baseline_test.py.

Checkpoints

Click here for checkpoints