Hi, first time writing in this forum.
Me and my colleague are using enobio32 and trying to run motor imagery CSP scenario for a project. We were using 10 sensors (FC5, FC1, FC2, FC6, CP5, CP1, CP2, CP6, C3, C4) and OpenViBE designer 1.1.0. When we got to online stage, it seemed working at first, but soon we realized that it was registering both left and right hand movements as right hand movements. Moreover the Graz visualization seemed to have been "balanced" not in the middle of the screen, but in the middle of the right hand side. It was responding, we just couldn't get it to move to the left side. During the classifier trainer stage we got only 58% performance, but because of limited time we went with it.
Was this the reason for our failure? Should we just try to get better performance rating and that should fix it? Generally we felt like there was little to adjust or alter in the scenario to make it work better for us, without understanding what exactly these boxes do.
Motor Imagery CSP example
-
- Posts: 775
- Joined: Tue Dec 04, 2012 3:53 pm
- Location: INRIA Rennes, FRANCE
Re: Motor Imagery CSP example
Hello mhlinka,
sorry for the late reply. I have personally experienced the same kind of problems with the OV motor imagery scenarios. The classifier accuracy from the cross-validation shouldn't be trusted too much due how to data is generated by the scenario ( see here : http://openvibe.inria.fr/tutorial-how-t ... te-better/ ) . In any case, as this reported percentage is *optimistic*, the number you report is too low to attain decent control, I would prefer at least 70%+ as a starting point. The OV scenarios are pretty much detecting ERD (event related desynchronization) in the beta band, a phenomenon which is not that clear in all subjects. You may need to either train people, or find a person who has natural talent. Classically in the papers the subjects have gone through hours of imagery training over a long period of time.
The skew of the classifier predictions as far as I currently understand is due to the nonstationarity, i.e. at least with some people, the recorded brain signal tends to change over time in its spectral characteristics. The current pipeline has hard time if this happens. Or, it could be a bug somewhere - if thats the case, I'd be delighted if somebody would tell where exactly and we'll patch it asap.
Best,
Jussi
sorry for the late reply. I have personally experienced the same kind of problems with the OV motor imagery scenarios. The classifier accuracy from the cross-validation shouldn't be trusted too much due how to data is generated by the scenario ( see here : http://openvibe.inria.fr/tutorial-how-t ... te-better/ ) . In any case, as this reported percentage is *optimistic*, the number you report is too low to attain decent control, I would prefer at least 70%+ as a starting point. The OV scenarios are pretty much detecting ERD (event related desynchronization) in the beta band, a phenomenon which is not that clear in all subjects. You may need to either train people, or find a person who has natural talent. Classically in the papers the subjects have gone through hours of imagery training over a long period of time.
The skew of the classifier predictions as far as I currently understand is due to the nonstationarity, i.e. at least with some people, the recorded brain signal tends to change over time in its spectral characteristics. The current pipeline has hard time if this happens. Or, it could be a bug somewhere - if thats the case, I'd be delighted if somebody would tell where exactly and we'll patch it asap.
Best,
Jussi
Re: Motor Imagery CSP example
would you provide a reference to these types of trainings?jtlindgren wrote: You may need to either train people, or find a person who has natural talent. Classically in the papers the subjects have gone through hours of imagery training over a long period of time.
-
- Posts: 775
- Joined: Tue Dec 04, 2012 3:53 pm
- Location: INRIA Rennes, FRANCE
Re: Motor Imagery CSP example
E.g.
"Graz-BCI: State of the Art and Clinical Applications", G. Pfurtscheller, C. Neuper, G. R. Müller, B. Obermaier, G. Krausz,
A. Schlögl, R. Scherer, B. Graimann, C. Keinrath, D. Skliris,, M. Wörtz, G. Supp, and C. Schrank,
IEEE Trans. Neural Systems and Rehabilitation Engineering, 2003.
"Control of a two-dimensional movement signal by a noninvasive brain–computer interface in humans",
Jonathan R. Wolpaw and Dennis J. McFarland, PNAS, 2004.
Cheers,
Jussi
"Graz-BCI: State of the Art and Clinical Applications", G. Pfurtscheller, C. Neuper, G. R. Müller, B. Obermaier, G. Krausz,
A. Schlögl, R. Scherer, B. Graimann, C. Keinrath, D. Skliris,, M. Wörtz, G. Supp, and C. Schrank,
IEEE Trans. Neural Systems and Rehabilitation Engineering, 2003.
"Control of a two-dimensional movement signal by a noninvasive brain–computer interface in humans",
Jonathan R. Wolpaw and Dennis J. McFarland, PNAS, 2004.
Cheers,
Jussi