EEGNAS: an EEG-targeted Neural architecture search algorithm

###Instructions to run cross/per subject experiment on the BCI Competition IV 2a dataset:

1. git clone the project to your machine
2. from the main folder 'BCI_Benchmarks' run: python EEGNAS_experiment.py -e cross_per_subject
3. This will run 2 experiment configurations:
   1. Cross subject directed NAS (initial data loading takes ~1 min)
   2. Per subject directed NAS

• Note: On a high-end Nvidia GPU each configuration should take 1-2 days to complete.
• For test purposes, it is possible to edit the 'config.ini' file, located in the 'configurations' folder.
  • For example, the property 'num_generations' can be changed from 75 to any other number, in order to get shorter run-times (and worse results).

###Instructions to run mixed data experiment on the BCI Competition IV 2a dataset:

• steps equivalent to the previous experiment, except stage 2:

2. from the main folder 'BCI_Benchmarks' run: python BCI_IV_2a_experiment.py -e cross_per_subject

Additional Notes

• Datasets:

  • The BCI Competition IV 2a dataset is included in the repository. No further action required.
  • BCI Competition IV 2b dataset - taken from the moabb BCI toolbox.
  • High Gamma dataset - need to download manually and move to the folder 'data/HG'
  • Inria BCI dataset - need to download from kaggle and move to the folder 'data/NER15'
  • Opportunity dataset - need to download manually and move to the folder 'data/Opportunity'. Use preproccesing code from this repository to obtain the file 'oppChallenge_gestures.data', which needs to be put in the folder 'data/Opportunity'

• The Moabb package is required to run EEGNAS, it isn't available on pip, but can be downloaded manually from git here: https://github.com/NeuroTechX/moabb

• EEGNAS is work in progress. In the future will be added an option to automatically analyze your own data and receive a neural architecture. Meanwhile, interested users can try to understand the NAS process themselves and add their own DB (the main task is to edit the file BCI_IV_2a_experiment.py and configurtions/config.ini)

• The repository is heavy because of the BCI Competition IV 2a dataset. It may take a while to download.

Citing

• If you use this code in a scientific publication, please cite us as:

@inproceedings{eegnas,
	title={EEGNAS : Neural Architecture Search for Electroencephalography Data Analysis and Decoding},
	author={Elad Rapaport, Oren Shriki, Rami Puzis},
	booktitle={Joint Workshop on Human Brain and Artificial Intelligence (HBAI)},
	year={2019},
	publisher={Springer}
}