[Insight-users] weighted Gaussian mixtures

[Daniel Mace]

Michel,

Have you tried SetWeights() on the GuassianMixtureModelComponent class?  
A quick pass on the code looks like it sets a prior for the class labels.

If that doesn't work, then the easiest solution would be just to 
replicate the data for that particular class (create 2,3 times the data 
points that are just replicates of the original data).  That *should* 
also help re-weight that class.

However, neither of those really seem to address the large variance of 
the class.  Have you tried plotting the histogram for the values of that 
class?  It may be just that you don't have many data points and the 
training of the model is getting skewed by a few outliers.  It could 
also be that your model choice specification is not appropriate for that 
particular class.

Cheers,
Dan

Audette, Michel wrote:
> Dear all, 
>
> I am working on expectation maximization of Gaussian mixtures, and data corresponding to one of the classes seem to be very flat, in relation to the other classes, so that EM tends to assign it a really large variance. I would like to use statistics computed from a small set of training points to not only initialize the expectation maximization, but provide  values which could be assigned a strong weight, so that they influence the final result more than in classic EM. 
>
> Is anyone aware of such as refinement to expectation maximization, and of a possible implementation in ITK? 
>
> Best regards, 
>
> Michel
>
> Michel Audette, Ph.D.          
> Innovation Center Computer Assisted Surgery (ICCAS)          
> Philipp-Rosenthal-Strasse 55         
> Leipzig, Germany         
> Phone: ++49 (0) 341 / 97 - 1 20 13         
> Fax: ++49 (0) 341 / 97 - 1 20 09         
>
> _______________________________________________
> Insight-users mailing list
> Insight-users at itk.org
> http://www.itk.org/mailman/listinfo/insight-users
>