ctorre/self-calibrating-bci

Self-Calibrating BCI Dataset (NeurIPS 2025)

Self-Calibrating BCIs: Ranking and Recovery of Mental Targets Without Labels

This dataset contains brain-computer interface (BCI) data from self-calibrating experiments where participants paid attention to faces while EEG signals were recorded. The dataset includes:

• Face representations in GAN latent space
• Processed EEG features from neural networks
• 9,234 experimental trials

📊 Dataset Summary

• Domain: Neuroscience, Brain-Computer Interfaces
• Task: Mental imagery, face recognition, BCI calibration
• Size: 9,234 samples, ~39 MB (HDF5 format)
• Format: HDF5
• License: MIT

🎯 Dataset Structure

The data is stored in HDF5 format under data with three main arrays:

Array	Shape	Dtype	Description
target_faces	(9234, 512)	float64	Target face latent vectors (Progressive GAN)
observed_faces	(9234, 512)	float64	Observed face latent vectors
eeg_net	(9234, 176)	float32	Neural network processed EEG features

All arrays are aligned: row i in each array corresponds to the same experimental trial.

Data Fields

target_faces

• Description: 512-dimensional latent vectors from Progressive GAN representing faces participants were trying to imagine
• GAN Model: Progressive GAN trained on CelebA-HQ 1024×1024
• Value Range: Approximately [-5, 5] (latent space coordinates)
• Usage: Ground truth for BCI task; can be decoded to face images using GAN decoder

observed_faces

• Description: 512-dimensional latent vectors for faces actually presented/selected during trials
• Relationship: Distance to target_faces measures BCI performance
• Usage: Compare with target_faces to evaluate mental imagery accuracy

eeg_net

• Description: 176-dimensional learned EEGNet representations from EEG signals
• Processing: Neural network feature extraction from EEG data
• Electrodes: Derived from 29-channel EEG system
• Usage: Input features for BCI decoding models

💻 Usage Examples

Run example_load_data.py.

📖 Data Collection

Experimental Setup:

• Participants imagined target faces while EEG was recorded
• 29-channel EEG system
• Face stimuli generated from Progressive GAN latent space
• Self-calibrating paradigm (no labeled training data)

Processing Pipeline:

1. Raw EEG → Windowing & feature extraction → 203 features
2. 203 features → Neural network → 176-dim embeddings (eeg_net)
3. Face images → GAN encoder → 512-dim latent vectors

📄 Citation

If you use this dataset, please cite:

@article{grizou2025self,
  title={Self-Calibrating BCIs: Ranking and Recovery of Mental Targets Without Labels},
  author={Grizou, Jonathan and de la Torre-Ortiz, Carlos and Ruotsalo, Tuukka},
  journal={Advances in Neural Information Processing Systems},
  year={2025}
}

📜 License

This dataset is released under the MIT License.

🔗 Related Resources

• Paper Repository: github.com/jgrizou/neurips-self-calibrating-bci
• GAN Checkpoint: Progressive GAN for CelebA-HQ 1024×1024
• Contact: [email protected]

🤝 Contributions

This dataset was created as part of the NeurIPS 2025 paper. For questions, issues, or suggestions, please contact [email protected] or open an issue on the paper repository.

🙏 Acknowledgments

Jonathan Grizou conducted this work during his tenure as an Assistant Professor at the University of Glasgow and subsequently through GrizAI Ltd. We gratefully acknowledge the financial support of both organizations.

This research was partially funded by the Alfred Kordelin Foundation (grant 230099) and the Finnish Foundation for Technology Promotion (grant 10168).

Computing and storage resources were provided by the Finnish Computing Competence Infrastructure (FCCI; HILE ERC grant ILLUMINATOR, 101114623).