How to create a classifier scenario?

[toncho11]

Hi,

I would like to test a classifier scenario. Please help.

1. How to first train the classifier for simple "Left"/"Right"?

2. How to use an already trained classifier?

-Anton

[yrenard]

Dear toncho11,

thank you for your interest in OpenViBE and welcome on this forum.

I'm glad that you got the emotiv build working all by yourself. In order to start with OpenViBE, I suggest that you look at the video tutorial, the classic HTML user tutorials and at the sample scenarios included in each release (check the share/openvibe-scenarios folder in the installation folder).

This will probably answer most of your first questions

Have fun with OpenViBE !
Yann

[toncho11]

I watched the video tutorial and specifically the classification part - I liked it!

I wonder if I could contribute with a "feature aggregator" and a "classifier processor". I have some experience with machine learning.

Can you please elaborate more on the:

1. What is the idea of the reference channel?

2. How is the spatial filter used to reduce signal complexity?

Thanks,
Anton

[ddvlamin]

1. What is the idea of the reference channel?

These can be the channels that are placed on "brain-activity-neutral" positions. The signals they measure are then assumed to be noise related and hence are substracted from the rest of the EEG channels. But there are different kind of reference systems such as bipolar which takes pair of electrodes that are substracted from each other. I'm not very sure what the exact purpose and meaning is of every montage (type of reference system used). Something that is sure is that phase information varies strongly between these montages and thus caution is required.

2. How is the spatial filter used to reduce signal complexity?

This depends on the type of spatial filter you use. Basically, it tries to transform the signal in a way that its main components capture or embed most of the available information in the signal. This is often done by some kind of optimization where one tries to maximize the signal to noise ratio. Off course, to do this, one has to be able to measure or quantify the noise part and the signal part. For example, for common spatial patterns (CSP), one uses signals of condition one (left hand motor imagery) as the "signal" part and signals of condition two (right hand motor imagery) as the "noise" part. Then it's just a matter of finding a vector that maximize this ratio and you have found a subspace that maximizes this signal to noise ratio. The xDAWN algorithm does something similar by estimating the P300 potentials and use them as "signal" part and the rest of the signal as noise part. By selecting only a few of these main components you can view it as some kind of dimensionality reduction where you retained most of the multivariate signal's information (but in fewer variables). As you have experience with machine learning, you will understand the link with PCA, which can also be seen as a spatial filter in this context.

These are spatial filters that needs to be computed from the data and use some kind of supervision, but there are also Laplacian filters that are completely unsupervised, simple yet effective and can be manually filled in in the spatial filter box.

I hope this helps a bit,
Best regards,
Dieter Devlaminck