Below are a few refs on incremental/ online EM. Some papers focus on
mixture models, but the theory can be easily generalized to any
directed graphical model.
HTH
Kevin
@incollection{Neal98,
author = "R. M. Neal and G. E. Hinton",
title = "A new view of the {EM} algorithm that justifies incremental
and other variants",
booktitle = "Learning in Graphical Models",
editor = "M. Jordan",
publisher = "MIT Press",
year = 1998,
}
@article{Cappe09,
title = {{Online EM Algorithm for Latent Data Models}},
author = "O. Cappe and E. Mouline",
year = 2009,
journal = jrssb,
volume = 71,
number = 3,
month = "June",
pages = "593--613"
}
@inproceedings{Liang09,
title = {{Online EM for Unsupervised Models}},
author = "P. Liang and D. Klein",
booktitle = "Proc. NAACL Conference"
}
On Thu, Mar 15, 2012 at 2:54 PM, Parot Ratnapinda <par...@gmail.com> wrote:
> Dear Colleagues,
>
>
>
> I am working on a system that performs classification tasks from a large
> continuous stream of data. This system has a fixed Bayesian network
> structure and applies an EM algorithm to update its' parameters. There are
> articles mention incremental learning of Bayesian network. However, my work
> does not involve any structure changes.
>
>
>
> My question concerns how I should perform incremental learning for
> parameters updating with a new data. Given that I have an existing network,
>
> How often should I do for incremental learning (ex. every 10, 100, or 1000).
> What is better in terms of accuracy for prediction: 1) learn incrementally
> with only new data 2) learn every time from the beginning (do batch learning
> at every k instances).
>
> Please feel free to give me any suggestions, comments, or point out to the
> existing works that I may have overlooked.
>
>
> Sincerely,
>
> Parot Ratnapinda
>
>
> _______________________________________________
> uai mailing list
> uai@ENGR.ORST.EDU
> https://secure.engr.oregonstate.edu/mailman/listinfo/uai
>
_______________________________________________
uai mailing list
uai@ENGR.ORST.EDU
https://secure.engr.oregonstate.edu/mailman/listinfo/uai
