locked
Model Evidence Calculation (Migrated from community.research.microsoft.com) RRS feed

  • Question

  • Vincent Tan posted on 06-17-2009 6:46 PM

    I tried to compute the model evidence for a Multiclass Probit Regression model. I find that sometimes the calculation converges to a sensible value but on several occasions it diverges. This is dependent on the number of iterations of the inference engine (I used EP). Is this to be expected because the precision matrices of the weights are nearly singular during the iterations? How can I get around this problem?

    Thanks, Vincent

    Friday, June 3, 2011 4:56 PM

Answers

  • C.Ranga Swamy replied on 06-20-2009 1:31 AM

    hi

              please give  me suggetions on  good basic mathmatics books and links related to approching of critical problems  solving with the basic mathamatics.

    Regards

    C.Ranga Swamy

     

     

    Friday, June 3, 2011 4:56 PM

All replies

  • C.Ranga Swamy replied on 06-20-2009 1:31 AM

    hi

              please give  me suggetions on  good basic mathmatics books and links related to approching of critical problems  solving with the basic mathamatics.

    Regards

    C.Ranga Swamy

     

     

    Friday, June 3, 2011 4:56 PM
  • minka replied on 07-03-2009 8:49 AM

    Hi Vincent, this is a known problem that we are looking at.  The issue is indeed the ill-conditioning of the precision matrices, and it shows up when EP is iterated too many times.  Limit the number of EP iterations and you should be fine.

    Friday, June 3, 2011 4:56 PM