B F M E E G 700m

Introduction

BFM-EEG-700M is a foundation model designed for processing time-series neuroimaging data, specifically electroencephalogram (EEG) signals. This model is based on the T5-large architecture and has been fine-tuned from the Chronos-t5-large model. It is developed by CERC-AAI and is suitable for various general-purpose brain signal processing tasks.

Architecture

The BFM-EEG-700M utilizes a T5-large model architecture, which is known for its versatility in handling a wide range of tasks. This architecture is particularly adapted for the processing of EEG data, making it a powerful tool for neuroimaging analysis. The model has tags indicating its broad applicability across general-purpose, brain, and signal processing domains.

Training

The model was trained using the NMT EEG dataset, an open-source collection of EEG recordings from both healthy and pathological subjects. This dataset includes 2,417 recordings, each with multichannel EEG data. The recordings are labeled based on the participants' pathological state, categorized as normal or abnormal. The EEG channels are treated as individual time series, divided into context and prediction target windows. The hyperparameters for training the model align with those specified in the Chronos paper.

Guide: Running Locally

To run the BFM-EEG-700M model locally, follow these steps:

1. Clone the Repository: Visit the CERC-AAI GitHub page and clone the repository.
2. Set Up Environment: Ensure you have the necessary dependencies installed, as specified in the repository's documentation.
3. Inference Notebook: Use the provided Colab notebook for inference on the MOABB dataset.
4. GPU Suggestion: Due to the model's size, it is recommended to use cloud GPUs such as those from AWS, Google Cloud, or Azure for efficient processing.

License

The model's license is currently unspecified. Please refer to the model card on Hugging Face or contact the developers at CERC-AAI for more information regarding usage and distribution rights.