We designed three python demos for SSVEP Classifier
- utils/cca_utils.py
- utils/vae_utils.py
- cca_ssvep.ipynb Classification Using Canonical Correaltion Analysis (CCA)
- cnn_ssvep.ipynb Classification Using Complex Spectrum Features and Convolutional Neural Networks (C-CNN)
- vae_ssvep.ipynb Classification Using Variational Autoencoder（VAE）and Convolutional Neural Networks (VAE-CNN)
Dataset Reference: 12-Class publicly available SSVEP EEG Dataset Dataset.
The following is implemented on a 12-Class publicly available SSVEP EEG Dataset
Dataset Download URL: https://github.com/mnakanishi/12JFPM_SSVEP/tree/master/data
Each .mat file has a four-way tensor electroencephalogram (EEG) data for each subject. Please see the reference paper for the detail.
[Number of targets, Number of channels, Number of sampling points, Number of trials] = size(eeg)
- cdot Number of targets : 12
- Number of channels : 8
- Number of sampling points : 1114
- Number of trials : 15
- Sampling rate [Hz] : 256
The order of the stimulus frequencies in the EEG data: [9.25, 11.25, 13.25, 9.75, 11.75, 13.75, 10.25, 12.25, 14.25, 10.75, 12.75, 14.75] Hz (e.g., eeg(1,:,:,:) and eeg(5,:,:,:) are the EEG data while a subject was gazing at the visual stimuli flickering at 9.25 Hz and 11.75Hz, respectively.)
The onset of visual stimulation is at 39th sample point, which means there are redundant data for 0.15 [s] before stimulus onset.
We tested the above three methods from EEG data of ten people.
| S1 | S2 | S3 | S4 | S5 | S6 | S7 | S8 | S9 | S10 | Overall | |
|---|---|---|---|---|---|---|---|---|---|---|---|
| CCA | 29.16% | 26.25% | 59.44% | 80.27% | 52.36% | 87.22% | 69.17% | 96.67% | 66.38% | 65.27% | 63.22% |
| C-CNN | 77.77% | 56.80% | 94.99% | 98.19% | 99.72% | 99.72% | 94.16% | 99.16% | 97.36% | 89.86% | 90.77% |
| VAE-CNN | 99.86% | 99.44% | 99.86% | 99.72% | 100% | 97.91% | 100% | 100% | 99.30% | 99.72% | 99.58% |
https://github.com/nikk-nikaznan/SSVEP-Neural-Generative-Models