locked
BPM vector gaussian (Migrated from community.research.microsoft.com) RRS feed

  • Question

  • anderson posted on 06-08-2010 12:12 PM

    I'm not sure I understand how the vector gaussian is created in the BPM example on your site.  What makes up the vector gaussian data in the BPM example?

    Friday, June 3, 2011 5:49 PM

Answers

  • minka replied on 06-25-2010 11:02 AM

    Yes, this VectorGaussian appears as wPosterior in the testing code.

    Friday, June 3, 2011 5:50 PM

All replies

  • minka replied on 06-11-2010 7:31 AM

    I assume you are talking about the multi-class BPM example.  In that example, the VectorGaussian objects represent prior distributions or posterior distributions on the weights.  When doing incremental training, we take the VectorGaussian posterior from one chunk and use it as the prior for the next chunk.

    Friday, June 3, 2011 5:50 PM
  • anderson replied on 06-14-2010 4:10 PM

    Sorry, I guess I should've been a bit more clear.  I was referring to this output from your Tutorial 4: Bayes Point Machine.

     

    Dist over w=
    VectorGaussian(0.0565 0.005873 -1.625, 0.0006731 -0.000166 -0.01058 )
    -0.000166 0.0001906 -0.002489
    -0.01058 -0.002489 0.3558

     

    So, the above are weights from data inferred by the training set.  Now, the test set will take the above weights to create its probabilities?

    Friday, June 3, 2011 5:50 PM
  • minka replied on 06-25-2010 11:02 AM

    Yes, this VectorGaussian appears as wPosterior in the testing code.

    Friday, June 3, 2011 5:50 PM