Online and offline testing

[stefan]

Hi everyone,

I’m do not quite understand what are the differences between the training scenario for P300 speller and the online scenarios? The layout looks very different, however, the differences on functionality are not clear to me. Also, I’d like to know if the accuracy showing at the end of training is considered offline testing or has a different meaning?

I replaced the letters with images for the speller, an error shows that no text to compare? What can be done for this?

Also, I used different channel sets using “channel selector” box, but the results were very close on the training, although I used only one channel some times and a list of 14 channels other times, however, the results were almost the same, which seems strange to me! Has anyone played with this box?

Any help is much appreciated
Cheers

[ddvlamin]

I’m do not quite understand what are the differences between the training scenario for P300 speller and the online scenarios? The layout looks very different, however, the differences on functionality are not clear to me.

In the online scenario one has to average the EEG for each row and column stimulus which can then be classified with the classifier as either a target row or column, this is not the case for the acquisition scenario. For now, it is my experience that it is best to use the online scenario for both acquisition and online use .

Also, I’d like to know if the accuracy showing at the end of training is considered offline testing or has a different meaning?

It is considered to be an estimation of how well the classifier will perform in online use (in case cross-validation is used). However as I explained in my previous post you will have to take into account the class imbalance of the data.

I replaced the letters with images for the speller, an error shows that no text to compare? What can be done for this?

Did you start from the dedicated P300 card scenario in the p300-magic-card directory?

Also, I used different channel sets using “channel selector” box, but the results were very close on the training, although I used only one channel some times and a list of 14 channels other times, however, the results were almost the same, which seems strange to me! Has anyone played with this box?

I only used the xDAWN algorithm with only one spatial filter and thus one virtual channel remaining which works quite well.

Best regards,
Dieter Devlaminck