Increasing classifier accuracy with many training data sets

Working with OpenViBE signal processing scenarios and doing scenario/BCI design
Post Reply
devash_h
Posts: 4
Joined: Fri Jun 09, 2017 2:19 pm

Increasing classifier accuracy with many training data sets

Post by devash_h »

Hi all,

I would like to know if there is a way of increasing the accuracy of the BCI example bundles.

I am using them as if and am getting very poor accuracy. Given the time period of my project, it won't be possible for me to extensively train people in Motor Imagery for the experiments.

I was wondering if there was a way of using multiple data sets to train the classifier. I currently have at least 15 different data sets from a group of people and I would wager that using all of them to train the classifier would significantly increase the accuracy.

Please do help me out asap if you can. I am running out of time.


Many thanks,
D

jtlindgren
Posts: 775
Joined: Tue Dec 04, 2012 3:53 pm
Location: INRIA Rennes, FRANCE

Re: Increasing classifier accuracy with many training data s

Post by jtlindgren »

Hi Devash,

technically that can be done simply by catenating the stuff from the different users before it goes to CSP trainer or the classifier trainer; in the first case you catenate signal files, in the second, feature vectors. One way to do this is to export these things to CSV files as they are before they enter the corresponding training box and then append the files after each other, remove the few extra lines (text editor works) and then load them back and feed to the trainer.

Unfortunately how much can be gained by this approach is a bit unclear. The reason is that the changes in the EEG patterns that motor imagery causes are quite different in different people, possibly due to anatomy, location of the motor cortices, what exactly is imagined, etc. These patterns (beta depression for example) might be differently asymmetric between different people and the left and right motor cortices, have different power and so need different weighting/thresholding from the classifier. Although there could be some transfer and similarity between people, as well as noise mitigation effects, if people 'cant do' motor imagery in a sense that difference between the left and right imagination is not really in the signal or its very weak (lost in noise), then there's not much signal processing or machine learning magic can do.

I don't mean that the approach could not improve things, but I have a hunch that you'd either need more clever classifiers than those in openvibe, and possibly even more data than 15 datasets, but could be worth a try. You'll get the catenation done in a few hours I imagine. Not sure how gracefully the openvibe memory use behaves in that case. An alternative is to build different classifiers and then vote on their outputs.

In any case, if you try some combination approach out, it'd be great to hear what results you obtained.

Cheers,
Jussi

devash_h
Posts: 4
Joined: Fri Jun 09, 2017 2:19 pm

Re: Increasing classifier accuracy with many training data s

Post by devash_h »

Dear Jussi,

That was extremely insightful and definitely gives me something to think about going forward.

I will definitely get back to you on my method and what results it yields.


Thank you very much for your time and help.


Best,
Devash

Post Reply