OpenViBE forum

Hello

I am trying to understand the working of the classifier trainer box with different strategies like: native, one vs all, one vs one etc.
In the case of a native classifier, the .cfg file consists of two sets of weights and bias values. Could someone please explain as to why we get two sets of values and what do they correspond to? I have added an example of the ssvep scenario below (classifier-1.cfg):

<OpenViBE-Classifier-Box XMLVersion="3">
<Strategy-Identifier class-id="(0xffffffff, 0xffffffff)">Native</Strategy-Identifier>
<Algorithm-Identifier class-id="(0x2ba17a3c, 0x1bd46d84)">Linear Discrimimant Analysis (LDA)</Algorithm-Identifier>
<Stimulations>
<Class-Stimulation class-id="1">OVTK_StimulationId_Label_01</Class-Stimulation>
<Class-Stimulation class-id="2">OVTK_StimulationId_Label_02</Class-Stimulation>
</Stimulations>
<OpenViBE-Classifier>
<LDA version="1">
<Classes>1 2 </Classes>
<Class-config-list>
<Class-config>
<Weights> 2.914286e+000 5.335379e+000 5.171503e+000 5.738135e+000 4.520307e+000 4.525663e+000</Weights>
<Bias>-10.7541</Bias>
</Class-config>
<Class-config>
<Weights> 2.450198e+000 2.511593e+000 2.759214e+000 4.343394e+000 2.603230e+000 2.174770e+000</Weights>
<Bias>-3.4796</Bias>
</Class-config>
</Class-config-list>
</LDA>
</OpenViBE-Classifier>
</OpenViBE-Classifier-Box>


Thank You
Regards
Aravind R

Hi

LDA in OpenVibe is multiclass, and classifies a vector X in class k = argmax_i { P(class=i | vector=X) }
There is one set of weights associated to each class to compute P(class=i | vector=X).

For more details: plugins/processing/classification/src/algorithms/ovpCAlgorithmClassifierLDA.cpp and plugins/processing/classification/src/algorithms/ovpCAlgorithmLDADiscriminantFunction.cpp

Best regards

Thank you so much!
Your clue on LDA implementation being multiple class, led me to PRML book by Bishop, where Chapter 4 explained a similar implementation.

I was initially looking at plugins/processing/classification/src/box-algorithms/ovpCBoxAlgorithmClassifierTrainer.cpp and processor.cpp. I will look into ovpCAlgorithmClassifierLDA.cpp and ovpCAlgorithmLDADiscriminantFunction.cpp, and will get back to you.

Something that we are trying to do is to use the same datasets and analyse on MATLAB so that our understanding of the algorithm gets better. We will try and study the code a bit more carefully and try to understand.

Thank You
Regards
Aravind R