Example using Quantum MDM on real data

[toncho11]

This PR is focused on testing the Quantum MDM on real data.
The process will take some time to validate the code and the results.

[gcattan]

Hi, very good. In addition, think we can test for different couples of distance and mean for both MDM and QuanticMDM.
I agree this will take some times. Let's try with Quantum=False in the first time as you suggested, and then do the validation with Quantum=True in a second PR.

Do you want me to take a look at the flake8 error?

[toncho11]

Can we disable it temporally for now? :)

Yes, there are a number of parameters to test:

• Covariance, ErpCovariance, XDawnCovariance
• number of filters for the XDawn Covariance
• ...

[toncho11]

List of things to test in order to evaluate the new algorithm:

1. WithinSubjectEvaluation -> done

2. Shrinkage(shrinkage=0.01) and Shrinkage(shrinkage=0.8) ?

3. In docplex.py,we can try also to change CobylaOptimizer(rhobeg=0.01, rhoend=0.0001) to CobylaOptimizer(rhobeg=0.1, rhoend=0.00001)

4. mean = logeuclid, distance = convex , check y_pred if it is all 0 or all 1s.

5. test merging all subjects and splitting in train/validate to see if the algorithm we generalize better on many subjects instead of per session.

[gcattan]

3. it has an importnace. Here with rho in the range 1...0.000001
   

[gcattan]

There is still some issue with the regularization of the convex mean, so you can only use mean=convex with distance=euclid.
It does help to change rho and shrinkage as follows:
rhobeg=2.1, rhoend=0.000001, shrinkage=0.9

We need to test with different datasets to see if this is consistent. Eventually, we could change the fit method inside QuanticMDM, to optimize these parameters depending on the data.

[gcattan]

The mean returned by the docplex model still returns some zeros. It is a problem for other distances than the euclidian one.

I tried the following:

• ignore NaN when getting the log of the matrix
• use a warm start
• replace zeros by some default values

It run, but the results were worst. @qbarthelemy do you have an idea?

[qbarthelemy]

I have no knowledge in the quantum field, but here are some comments.

[gcattan]

Here are some of the results.

mean=convex/dist=convex run without error with Xdawn covariances (with corrected estimators as suggested above). The results are not good, but also it is not the same space.

In blue is a VotingClassifier with mean=convex/dist=euclid and mean=logeuclid/dist=convex.