[Topic-models] Expressive power of latent variable models

Alta de Waal alta.dewaal at gmail.com
Wed Jun 7 06:19:05 EDT 2017


Dear Topic-models digest,

I read in Murphy (Machine Learning - a probabilistic perspective) that
finite mixture models are limited in expressive power and therefore
continuous latent variable models are suggested for topic models.

I have a few questions on this statement:
1. Why is it limited in expressive power?
2. Can we measure expressive power of a model?
3. Is it true for all types of discrete data. For example, will it be true
for short text (where length of document plays a role?)

I would appreciate any referral to research papers on this topic as well.

Kind regards,
Alta de Waal
Department of Statistics
University of Pretoria
South Africa
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.cs.princeton.edu/pipermail/topic-models/attachments/20170607/e08ec173/attachment.html>


More information about the Topic-models mailing list