The 4th place solution for the Cortana competition organized on the Microsoft Azure ML platform.
Despite its good performance in the competition, this was a highly exploratory work which lacks some structure and good coding style.
ECoG time series of 800 ms windows centered at the visual stimulus presentation: pictures of houses or faces. The task is to identify the category of presented picture from brain signals.
The general pipeline is the following:
- learn a bunch of logistic regression classifiers with different combination of features
- select the best models based on 5-fold cross-validation
- compute a weighted average of first-layer model predictions or stack them with SVM
Features:
- Event-related potential (ERP)
- Event-related broadband (ERBB)
- Band powers for 4-10 Hz estimated with wavelet transform
- Covariance matrix projected on Riemannian tangent space
Warning Training may take up to 1 hour.
After training the models with train.py, they will be saved in the bundle folder:
- Zip the bundle folder and upload it to Azure platform as a dataset.
- Create an experiment with the structure shown below.
- Copy-paste one of the classifying scripts into the 'Execute Python Script' module.
