Hello,
I'm using the P300-Speller-xDAWN scenario, but I want to change the classifier of this scenario.
The default classifier is the LDA, but I need to use the MLP classifier, but when I change it, the online scenario does not work anymore.
Someone has used the MLP classifier in this scenario?
Thank you in advance.
How I change the classifier of P300-Speller-xDAWN scenario?
-
willianmrs
- Posts: 3
- Joined: Wed May 04, 2016 6:12 pm
-
jtlindgren
- Posts: 775
- Joined: Tue Dec 04, 2012 3:53 pm
- Location: INRIA Rennes, FRANCE
Re: How I change the classifier of P300-Speller-xDAWN scena
Hi, whats exactly the problem?
Some ideas in any case,
In OV 1.2.0 there's a small glitch in some demo scenarios related to the classifier trainer box. This bug can be sidestepped by replacing the classifier trainer box in the scenario with a new one and configuring it (we'll post a bugfix release later).
Secondly, the P300 design is very unbalanced, there's many more labels in the out-class than in the in-class, and subsequently the MLP often seems to learn a model that predicts the majority class. You can try to address this a little by toggling 'balance labels' setting 'on' or perhaps by tweaking the MLP training parameters. The spelling will definitely not work correctly if the cross-validation claims 0% accuracy for the other class.
However, that might not a coding bug but just a result of the backprop training algorithm on label-skewed and highly noisy data.
Good luck,
Jussi
Some ideas in any case,
In OV 1.2.0 there's a small glitch in some demo scenarios related to the classifier trainer box. This bug can be sidestepped by replacing the classifier trainer box in the scenario with a new one and configuring it (we'll post a bugfix release later).
Secondly, the P300 design is very unbalanced, there's many more labels in the out-class than in the in-class, and subsequently the MLP often seems to learn a model that predicts the majority class. You can try to address this a little by toggling 'balance labels' setting 'on' or perhaps by tweaking the MLP training parameters. The spelling will definitely not work correctly if the cross-validation claims 0% accuracy for the other class.
However, that might not a coding bug but just a result of the backprop training algorithm on label-skewed and highly noisy data.Good luck,
Jussi