These are chat archives for bayespy/bayespy

4th
Apr 2016
Jaakko Luttinen
@jluttine
Apr 04 2016 05:15
@deebuls about the first one: there is non-conjugacy in beta w.r.t. theta. also, what does 5xDirichlet mean? multiply the prob.vector by 5? i don't understand why wouldn't you just learn beta freely. why constrain it to be a probability vector?
@deebuls about the second model: again, non-conjugacy in beta w.r.t. theta. i don't quite understand the model construction. would it be ok to learn beta as a free parameter and discard alpha and phi?
Jaakko Luttinen
@jluttine
Apr 04 2016 06:26
@deebuls in develop branch, there's a node called DirichletConcentration. it can be used to find maximum likelihood estimate for the dirichlet concentration in a hierarchical model.
hmm.. documentation is poor, sorry. gotta fix. DirichletConcentration takes one mandatory argument: dimensionality of the probability vector space. that is, one integer.
Jaakko Luttinen
@jluttine
Apr 04 2016 12:47
@deebuls about the books: i started collecting some pieces from here and there to form a vb book, but it's really just a collection of copy-pasted random parts, missing lots of stuff. but just in case you find it in any way useful: http://variational-bayes-book.readthedocs.org/en/latest/
but as i said, it's very much a work in progress
Jaakko Luttinen
@jluttine
Apr 04 2016 12:53
but yeah, probably that is in no way useful atm :D
Deebul Nair
@deebuls
Apr 04 2016 15:54
@jluttine Thanks for the book. I will read it to get details of VB inference . But currently I am more interested in application of graphical modelling . To understand different models available when to use, different distributions when to use and where to use .
@jluttine regarding the questions . The models are actually proposed in https://probmods.org/hierarchical-models.html .
The explanation given is basically "sharing of knowledge" . So using these model we can learn about other bags content just by observing 1 bag . As all bags will be connected. And the learning is faster . They call the concept as Over hypotheses a part of Hierarchical modelling. The base paper is http://web.mit.edu/cocosci/Papers/KempPerforsTenenbaum06.pdf
Deebul Nair
@deebuls
Apr 04 2016 15:59
I dont know the details but like to implement it and know how it works . If you could suggest me some models I can try them also. I am on a learning curve .
Regarding Dirichletconcentration .. I didnt get where to use it . Can you be more elaborate on where I should use it . Or suggest some reading topics what it is :worried: . Sorry for the trouble
Deebul Nair
@deebuls
Apr 04 2016 16:14
Finally I have experimented with Mixture modelling on the same urn ball example. I am facing some problems . I have raised a issue #63 . Please have a look . Thanks