Hi,
I'm using the motor imagery scenario with csp. I have connected it to a game so the user can move two boxes by imagining right or left hand movements.
As I have noticed, in the 4th scenario(the online scenario) we don't need to use a cue; since when we train the LDA we give 1 second long epochs of the 4 second long training example every 1/16 second. although in the 4th scenario a Graz visualizer is used which gives the user cues .
here are my questions :
why we don't use the same "1 sec long every 1/16 sec" epochs to train the CSP filter ? instead the scenario sends the whole 4 second epoch to the CSP classifier as one training example.
Can I use the online scenario without cues or do I actually need to send cues from my game to openvibe to tell the user which direction of hand movement to think about and when?
Cheers,
Kiyarash
Should a cue be used in motor imagery scenario with csp
-
jtlindgren
- Posts: 775
- Joined: Tue Dec 04, 2012 3:53 pm
- Location: INRIA Rennes, FRANCE
Re: Should a cue be used in motor imagery scenario with csp
Hi Kiyarash,
thats a great question about CSP, especially since I don't have a good answer. From a textbook machine learning viewpoint, the overlapping epochs approach is suspicious since it makes the feature vectors quite clearly less "I.I.D." (independent and identically distributed). On the other hand, it potentially gives the classifier/csp minor variations around each epoch. Perhaps this has some utility. Of course the main point of having a sliding window over the data is that you get control commands more often and not only after each trial. This may allow more fluid control; given that it is accurate.
If you like, you can add similar epoching after CSP and try to compare the results (but do *not* look at the classifier trainer xval results, those will get just more biased towards optimism).
To classify data for your game you of course don't need cues, you just need to send data to your classification pipeline in a streaming fashion and decide what your game does with the predictions.
Cheers,
Jussi
thats a great question about CSP, especially since I don't have a good answer. From a textbook machine learning viewpoint, the overlapping epochs approach is suspicious since it makes the feature vectors quite clearly less "I.I.D." (independent and identically distributed). On the other hand, it potentially gives the classifier/csp minor variations around each epoch. Perhaps this has some utility. Of course the main point of having a sliding window over the data is that you get control commands more often and not only after each trial. This may allow more fluid control; given that it is accurate.
If you like, you can add similar epoching after CSP and try to compare the results (but do *not* look at the classifier trainer xval results, those will get just more biased towards optimism).
To classify data for your game you of course don't need cues, you just need to send data to your classification pipeline in a streaming fashion and decide what your game does with the predictions.
Cheers,
Jussi