On Thu, May 8, 2014 at 4:57 PM, Jesse Hersch <[email protected]> wrote:
> On Thursday, May 8, 2014 4:14:32 PM UTC-7, William wrote:
>>
>> I could be wrong, but I don't think the implementation of Baum-Welch
>> is wrong.  The BM algorithm [1] using double precision numbers (which
>> is all the HMM algorithm in Sage uses) can lead to overflow, given the
>> sort of computations that are involved.
>
>
> Thanks for the reply!
>
> My understanding is that it's underflow that's more common with HMM stuff,
> due to all the products of small probabilities running around.  In some
> implementations I've seen this handled by the logsumexp trick:
> http://machineintelligence.tumblr.com/post/4998477107/the-log-sum-exp-trick
>
> Also in the Rabiner tutorial there's a section on scaling where he talks
> about underflow and how to handle it.  that's on page 16 (272) here:
> http://people.sabanciuniv.edu/berrin/cs512/reading/rabiner-tutorial-on-hmm.pdf
>
> Do you recall if you handled the underflow problem in your implementation?

I believe it does not.

> I haven't studied the code yet, but it seems like this could be the culprit.

I think you're right.  You should implement it!

-- William

>
> --
> You received this message because you are subscribed to the Google Groups
> "sage-support" group.
> To unsubscribe from this group and stop receiving emails from it, send an
> email to [email protected].
> To post to this group, send email to [email protected].
> Visit this group at http://groups.google.com/group/sage-support.
> For more options, visit https://groups.google.com/d/optout.



-- 
William Stein
Professor of Mathematics
University of Washington
http://wstein.org

-- 
You received this message because you are subscribed to the Google Groups 
"sage-support" group.
To unsubscribe from this group and stop receiving emails from it, send an email 
to [email protected].
To post to this group, send email to [email protected].
Visit this group at http://groups.google.com/group/sage-support.
For more options, visit https://groups.google.com/d/optout.

Reply via email to