Emaad Manzoor created MAHOUT-1484:
-------------------------------------
Summary: Spectral algorithm for HMMs
Key: MAHOUT-1484
URL: https://issues.apache.org/jira/browse/MAHOUT-1484
Project: Mahout
Issue Type: New Feature
Reporter: Emaad Manzoor
Priority: Minor
Following up with this
[comment|https://issues.apache.org/jira/browse/MAHOUT-396?focusedCommentId=12898284&page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#comment-12898284]
by [~isabel] on the sequential HMM
[proposal|https://issues.apache.org/jira/browse/MAHOUT-396], is there any
interest in a spectral algorithm as described in: "A spectral algorithm for
learning hidden Markov models (D. Hsu, S. Kakade, T. Zhang)"?
I would like to take up this effort.
This will enable learning the parameters of and making predictions with a HMM
in a single step. At its core, the algorithm involves computing estimates from
triples of observations, performing an SVD and then some matrix multiplications.
This could also form the base for an implementation of "Hilbert Space
Embeddings of Hidden Markov Models (L. Song, B. Boots, S. Saddiqi, G. Gordon,
A. Smola)".
--
This message was sent by Atlassian JIRA
(v6.2#6252)
