Page 1 of 1

Epoch duration

Posted: Sat Oct 03, 2020 10:22 am
by etc
Sorry for my bad English!
In this example:
Flash duration: 0.1
No flash duration: 0.5
Epoch duration: 1.2 sec.,
I think that a target-epoch (i.e. with the start from the stimulation target), will get the EEGs for one target stimulation and also for one no-target stimolation.
Is this a problem for the Classifier trainer?
Perhaps, is it better to choose an Epoch duration that put in every Epoch only one type of epoch? In the example, with Epoch duration 0.6, I think that in every epoch I will put the EEG of only one type.
Or is it not a problem?
Thank you!
Ettore

Re: Epoch duration

Posted: Tue Oct 06, 2020 10:21 am
by Thibaut
Hi,
I need more background please.^^
I think you are in P300 style if I remember good.
Share the scenario and more information.
Thibaut

Re: Epoch duration

Posted: Sun Oct 11, 2020 9:33 am
by etc
Thank you very much Thibaut for your answer!
Sorry for the delay of my new answer: I had the fever!
You remember very well! I am tryng to work with the P300 Speller scenario.
The question is:
if (for i.e.):
1) in the box "P300 stimulator" (in the P300 Speller acquisition) I indicate a flash duration of 0,2 ms. and a no flash duration of 0,1 ms.;
2) in the box "Stimulation based epoching" ( in the P300 Speller "train classifier") I indicate an epoch duration long 0,9 ms.;
Then I think that I will get epochs overlapping.
Now the epochs overlapping are a problem for the Classifier trainer? I think that with the overlapping the Classifier trainer will get in input epochs started with target stimulus (and, then, classified in input as target) with inside data "after" stimulus target but also data "after" no target.
Is it a problem? Is it better to indicate an epoch duration equal to the sum of flash duration and no flash duration to get for every epoch only one type of data?
Sorry, I have a second question: at the end of the work of the train classifier we get 2 confusion matrix.
I think to have understood the first matrix ("cross validation test accuracy") because I think to have understood as the cross-validations works.
I didn't understand the second matrix. I would ask to you if you can spend some words about the "training set accuracy" of the second matrix. (I got 61% on cross validation test accuracy and 100% on training set accuracy).
Thank you very much!
Ettore

Re: Epoch duration

Posted: Thu Jan 07, 2021 11:19 am
by fabien.lotte
Dear Ettore,

Sorry, I saw that we have not answered yet your questions, sorry for the delay.
Regarding the overlap in epochs, indeed, that is the case, in a given epoch there might be presence of ERP to subsequent stimulus. However, that is taken into account when training the XDAWN spatial filter, so that is OK to do so. Moreover, using shorter epochs (to avoid these overlaps) would mean too short epochs to properly detect the P300.

Regarding the two confusion matrices, the "training accuracy" is the matrix representing the accuracy obtained when applying the classifier on the same data used to train this classifier. It is thus very biased, and not very representative of the real performances (which is why you got 100%). The cross validation accuracy applies a classifier on an unseen part of the training data, thus giving a better estimate of the performances.

I hope this helps,

Re: Epoch duration

Posted: Thu Jan 07, 2021 1:28 pm
by etc
Thank you very much!!!
Ettore