Detail publikačního výsledku

Restricted Boltzman Machines for Image Tag Suggestion

KRÁL, J.; HRADIŠ, M.

Originální název

Restricted Boltzman Machines for Image Tag Suggestion

Anglický název

Restricted Boltzman Machines for Image Tag Suggestion

Druh

Stať ve sborníku mimo WoS a Scopus

Originální abstrakt

In this paper, we propose to model dependencies among binary variables in semantic tagging and similar tasks by Restricted Boltzmann Machines (RBM). In the proposed approach, Gibbs sampling allows learning RBMs even on data with large portion of missing values. Similarly, Gibbs sampling is used to estimate marginal probabilities of tags. The results show that the tag predictions become more certain with higher portion of known tags, and that the approach could be used for tag suggestion or semi-supervised learning.

Anglický abstrakt

In this paper, we propose to model dependencies among binary variables in semantic tagging and similar tasks by Restricted Boltzmann Machines (RBM). In the proposed approach, Gibbs sampling allows learning RBMs even on data with large portion of missing values. Similarly, Gibbs sampling is used to estimate marginal probabilities of tags. The results show that the tag predictions become more certain with higher portion of known tags, and that the approach could be used for tag suggestion or semi-supervised learning.

Autoři

KRÁL, J.; HRADIŠ, M.

Vydáno

26.04.2012

Nakladatel

Brno University of Technology

Místo

Brno

Kniha

Proceedings of the 19th Conference STUDENT EEICT 2012

Strany od

1

Strany do

5

Strany počet

5

URL

http://www.feec.vutbr.cz/EEICT/2012/sbornik/03doktorskeprojekty/09grafikaamultimedia/03-ikral.pdf

BibTex

@inproceedings{BUT192815,
  author="Jiří {Král} and Michal {Hradiš}",
  title="Restricted Boltzman Machines for Image Tag Suggestion",
  booktitle="Proceedings of the 19th Conference STUDENT EEICT 2012",
  year="2012",
  pages="1--5",
  publisher="Brno University of Technology",
  address="Brno",
  url="http://www.feec.vutbr.cz/EEICT/2012/sbornik/03doktorskeprojekty/09grafikaamultimedia/03-ikral.pdf"
}