Neuromorphic EEG classification for energy efficient brain-computer interfaces

Code for creating a deep learning based spiking EEG classifier, accompanying: https://ieeexplore.ieee.org/abstract/document/9629621

In order to to run the whole pipeline:

1. Use train_CNNs.ipynb to train CNNs on Graz IV 2b (http://www.bbci.de/competition/iv/, load with moabb)
2. Use convert_all.py to convert the CNNs to their spiking equivalent
3. Use runSCNN.py to run the spiking neural networks

Note: if you don't have access to a SpiNNaker you can run the model in Nest

The code is tested with python 3.8.3 and the following package versions:

tensorflow.keras == 2.4.0
tensorflow == 2.3.0
scikit-learn == 0.23.1
scipy == 1.6.1
numpy == 1.16.1
opencv-python == 4.5.1.48
mne == 0.22.0 (https://mne.tools/stable/index.html)
moabb == 0.2.1 (https://github.com/NeuroTechX/moabb)
braindecode == 0.5 (https://github.com/braindecode/braindecode)
pynn == 0.9.2 (https://neuralensemble.org/PyNN/)
pynest == 2.20.0 (https://www.nest-simulator.org/)
snntoolbox == 0.5.0 (https://snntoolbox.readthedocs.io/en/latest/guide/installation.html) \

For snntoolbox, if you want to use SpatialDropout2D, you have to add this to the layers to be ignored in the conversion process (on snntoolbox/parsing/utils.py, line 450).

When using the SpiNNaker chip, if you get a decrease in accuracy due to lost packages, you can try to increase the spikes_per_second parameter in the SpiNNaker config file.