Subject-Invariant-SSVEP-GAN

Code to accompany our International Conference on Pattern Recognition (ICPR) paper entitled - Leveraging Synthetic Subject Invariant EEG Signals for Zero Calibration BCI.

Code Structure

Main Components

• generate_sisgan.py - Our proposed SIS-GAN model for generating subject-invariant SSVEP-based EEG data
• train_subject_classifier.py - Subject classification network for biometric pattern detection
• ssvep_classification.py - SSVEP frequency classification network
• plot_subject_softmax.py - Subject-invariance evaluation using softmax probability analysis
• models.py - Neural network architectures for all models

Supporting Modules

• utils.py - Helper functions for data loading, logging, and model utilities

Dependencies and Requirements

1. Create a new uv environment with Python 3.12:

uv venv --python 3.12

2. Install the required dependencies defined in pyproject.toml:

uv pip install -e .

How to Use

Please follow the steps below to run the code to train the various components. Please read the Configuration section to understand how to modify the training parameters to suit real-world applications. The number of epochs is set to 1 by default for demonstration purposes, but you will need to increase this for meaningful results.

Complete Zero-Calibration Pipeline (Leave-One-Out)

This is the main pipeline for zero-calibration BCI evaluation, where we simulate having an unseen subject:

Step 1: Create pretrained subject classification weights

uv run python -m sis_gan.train_subject_classifier --validation_strategy loo

This trains subject classifiers using leave-one-out cross-validation, creating pretrained weights for each "unseen" subject scenario.

Step 2: Generate synthetic EEG data

uv run python -m sis_gan.generate_sisgan

Trains the SIS-GAN using data from "seen" subjects (excluding the test subject) and generates synthetic SSVEP signals.

Step 3: Evaluate SSVEP classification performance

uv run python -m sis_gan.ssvep_classification

Tests how well SSVEP frequency classification works when trained on synthetic data and tested on real unseen subject data.

Step 4: Verify subject-invariance

uv run python -m sis_gan.plot_subject_softmax

Analyzes whether the generated synthetic data contains subject-specific biometric patterns. Low, uniformly distributed probabilities indicate successful subject-invariance.

Subject Classification Only (Stratified Split)

If you only want to evaluate subject classification performance on the full dataset without running the complete pipeline:

uv run python -m sis_gan.train_subject_classifier --validation_strategy stratified

Performs standard train/test split (80/20) for subject classification evaluation across the entire dataset.

Data

The sample_data folder contains randomly generated data representing the shape of the input EEG data. Note: This is not real EEG data and is provided only for demonstration purposes.

Configuration

Model configurations are controlled using YAML files in the config directory. Modify these files to customize:

• Network architectures
• Training hyperparameters
• Data processing parameters
• Validation strategies

Note: By default, all models are only trained for a single epoch for demonstration purposes. Adjust the num_epochs parameter in the config files to train for longer - you will need to increase the number of epochs for meaningful results.

Development

Run pre-commit hooks

uv pip install -e ".[dev]"
uv run pre-commit install
uv run pre-commit run --all-files

Key Concepts

• Subject-Invariant: Generated data should not contain biometric patterns specific to individual subjects
• Zero-Calibration: Using synthetic data to train models for completely unseen subjects without any calibration data
• Leave-One-Out: Each subject is treated as "unseen" while others are used for training, simulating real-world deployment scenarios

Citation

Please cite the associated paper if you use this code:

@inproceedings{aznan2021leveraging,
  title={Leveraging Synthetic Subject Invariant EEG Signals for Zero Calibration BCI},
  author={Aznan, Nik Khadijah Nik and Atapour-Abarghouei, Amir and Bonner, Stephen and Connolly, Jason D and Breckon, Toby P},
  booktitle={2020 25th International Conference on Pattern Recognition (ICPR)},
  pages={10418--10425},
  year={2021},
  organization={IEEE}
}