[Topic-models] Why is E step of LDA iterative

Veena T veenat2005 at gmail.com
Sun Nov 23 22:59:40 EST 2008


Yes, even I am facing the sam eproblem. Say when I have some K number of
\phi values, according to theory they should add upto one. They are no doubt
addiing upto one. However, as you said, only one of the K values are
becoming one and rest all are zero. Even I am worried with this problem.
Anyone else has an answer for this?

On Sun, Nov 23, 2008 at 12:44 PM, Yuanlong Shao <shaoyuanlong at cad.zju.edu.cn
> wrote:

>  Sorry, I mean how to ensure the convergence and the correctness of
> results when converged.
>
> In my problem, where different configuration is used, some variables
> similar to \phi in LDA have
>
> non-maximum suppression like effects, which means that only the max values
> in the initial \phi
>
> get the value 1 and others becomes 0, but I don't know why.
>
>
>
> Yuanlong Shao
>
>
>
> *From:* Veena T [mailto:veenat2005 at gmail.com]
> *Sent:* 2008年11月23日 14:57
> *To:* Yuanlong Shao
>
> *Subject:* Re: [Topic-models] Why is E step of LDA iterative
>
>
>
> Generally we ensure convergence by looking into the likelihood value. if it
> further does not change, you can stop.
>
> 2008/11/23 Yuanlong Shao <shaoyuanlong at cad.zju.edu.cn>
>
> Thank you. But I still don't understand, normally, when max f(x,y), we
> compute partial derivatives
>
> for each variables and set them to 0, we get an equation set and solve
> them,
>
> but for this problem, since closed form solution does not exist, we are
> iteratively updating them,
>
> then how to ensure the convergence? I'm trying to figure whether I can
> apply the same method
>
> to similar problems.
>
>
>
> Best Regards,
>
>
>
> Yuanlong Shao
>
> State Key Lab of CAD&CG
>
> Zhejiang University
>
>
>
> *From:* Veena T [mailto:veenat2005 at gmail.com]
> *Sent:* 2008年11月23日 14:17
> *To:* shaoyuanlong at cad.zju.edu.cn
> *Cc:* topic-models at lists.cs.princeton.edu
> *Subject:* Re: [Topic-models] Why is E step of LDA iterative
>
>
>
> >Hi, all
>
>
>
> >Why is the E step of LDA iterative? I mean the updating formula of \gamma
> >and \phi.
> Because there is interdependency between themselves
>
>
>
> >In my algorithm, there are more variables to be updated in E step,
>
> >is there any situation that such iterative E step does not converge or
> >oscillate repeatedly?
> I am not sure
>
>
>
> Shao
>
>
> --
> Veena Srinivas,
> PhD scholar,
> Speech and Vision Lab
> IIT Madras, Chennai-36
> Tamil Nadu
>
>
>
>
> --
> Veena Srinivas,
> PhD scholar,
> Speech and Vision Lab
> IIT Madras, Chennai-36
> Tamil Nadu
>



-- 
Veena Srinivas,
PhD scholar,
Speech and Vision Lab
IIT Madras, Chennai-36
Tamil Nadu
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.cs.princeton.edu/pipermail/topic-models/attachments/20081124/155ec074/attachment.htm>


More information about the Topic-models mailing list