File Download
  Links for fulltext
     (May Require Subscription)
Supplementary

Conference Paper: Entropy coding for training deep belief networks with imbalanced and unlabeled data

TitleEntropy coding for training deep belief networks with imbalanced and unlabeled data
Authors
KeywordsPhysics
Sound
Issue Date2012
PublisherAcoustical Society of America. The Journal's web site is located at http://asa.aip.org/jasa.html
Citation
The ACOUSTICS 2012 Hong Kong Conference & Exihibition, Hong Kong, 13-18 May 2012. In Journal of the Acoustical Society of America, 2012, v. 131 n. 4, p. 3235, abstract no. 1aSCb1 How to Cite?
AbstractTraining deep belief networks (DBNs) is normally done with large data sets. In this work, the goal is to predict traces of the surface of the tongue in ultrasoundimages of the mouth during speech. Performance on this task can be dramatically enhanced by pre-training a DBN jointly on human-supplied traces and ultrasoundimages, then training a modified version of the network to predict traces from ultrasound only. However, hand-tracing the entire dataset of ultrasoundimages is extremely labor intensive. Moreover, the dataset is highly imbalanced since many images are extremely similar. This work presents a bootstrapping method which takes advantage of this imbalance, iteratively selecting a small subset of images to be hand-traced, then (re)training the DBN, making use of an entropy-based diversity measure for the initial selection. With this approach, a three-fold reduction in human time required to trace an entire dataset with human-level accuracy was achieved.
DescriptionSession 1aSCb - Speech Communication: Speech Processing Potpourri (Poster Session): no. 1aSCb1
Persistent Identifierhttp://hdl.handle.net/10722/211020
ISSN
2015 Impact Factor: 1.572
2015 SCImago Journal Rankings: 0.938

 

DC FieldValueLanguage
dc.contributor.authorBerry, J-
dc.contributor.authorFasel, I-
dc.contributor.authorFadiga, L-
dc.contributor.authorArchangeli, D-
dc.date.accessioned2015-06-30T07:55:39Z-
dc.date.available2015-06-30T07:55:39Z-
dc.date.issued2012-
dc.identifier.citationThe ACOUSTICS 2012 Hong Kong Conference & Exihibition, Hong Kong, 13-18 May 2012. In Journal of the Acoustical Society of America, 2012, v. 131 n. 4, p. 3235, abstract no. 1aSCb1-
dc.identifier.issn0001-4966-
dc.identifier.urihttp://hdl.handle.net/10722/211020-
dc.descriptionSession 1aSCb - Speech Communication: Speech Processing Potpourri (Poster Session): no. 1aSCb1-
dc.description.abstractTraining deep belief networks (DBNs) is normally done with large data sets. In this work, the goal is to predict traces of the surface of the tongue in ultrasoundimages of the mouth during speech. Performance on this task can be dramatically enhanced by pre-training a DBN jointly on human-supplied traces and ultrasoundimages, then training a modified version of the network to predict traces from ultrasound only. However, hand-tracing the entire dataset of ultrasoundimages is extremely labor intensive. Moreover, the dataset is highly imbalanced since many images are extremely similar. This work presents a bootstrapping method which takes advantage of this imbalance, iteratively selecting a small subset of images to be hand-traced, then (re)training the DBN, making use of an entropy-based diversity measure for the initial selection. With this approach, a three-fold reduction in human time required to trace an entire dataset with human-level accuracy was achieved.-
dc.languageeng-
dc.publisherAcoustical Society of America. The Journal's web site is located at http://asa.aip.org/jasa.html-
dc.relation.ispartofJournal of the Acoustical Society of America-
dc.rightsJournal of the Acoustical Society of America. Copyright © Acoustical Society of America.-
dc.rightsCreative Commons: Attribution 3.0 Hong Kong License-
dc.subjectPhysics-
dc.subjectSound-
dc.titleEntropy coding for training deep belief networks with imbalanced and unlabeled data-
dc.typeConference_Paper-
dc.identifier.emailArchangeli, D: darchang@hku.hk-
dc.identifier.authorityArchangeli, D=rp01748-
dc.description.naturepublished_or_final_version-
dc.identifier.doi10.1121/1.4708066-
dc.identifier.volume131-
dc.identifier.issue4-
dc.identifier.spage3235, abstract no. 1aSCb1-
dc.identifier.epage3235, abstract no. 1aSCb1-
dc.publisher.placeUnited States-

Export via OAI-PMH Interface in XML Formats


OR


Export to Other Non-XML Formats