ML Anthology

Expectation Maximization for Weakly Labeled Data

Yuri A. Ivanov, Bruce Blumberg, Alex Pentland
ICML 2001 pp. 218-225
/icml/2001/ivanov2001icml-expectation/

Abstract

We call data weakly labeled if it has no exact label but rather a numerical indication of correctness o the label "guessed" by the learning algo rithm - a situatio n cofi;)?v enco( tered inpro?(:fi o reinfo rcement learning. The term emphasizes similarities o o r approfi h to the kno wn techniqueso fso lving unsupervised and transductive pro blems. In this paper we present an on-line algoj thm that casts the pro blem as a multi-arm bandit with hidden state and so lves it iteratively within the Expectatio n-Maximizatio framewo7fi The hidden state is represented by a parameterized pro bability distributio no ver states tied to the reward. The parameterizatio n isfo rmally justified, allowing for smooth blending between likeliho o d- and reward-based costs.

Semantic Scholar

Cite

Text

Ivanov et al. "Expectation Maximization for Weakly Labeled Data." International Conference on Machine Learning, 2001.

Markdown

[Ivanov et al. "Expectation Maximization for Weakly Labeled Data." International Conference on Machine Learning, 2001.](https://mlanthology.org/icml/2001/ivanov2001icml-expectation/)

BibTeX

@inproceedings{ivanov2001icml-expectation,
  title     = {{Expectation Maximization for Weakly Labeled Data}},
  author    = {Ivanov, Yuri A. and Blumberg, Bruce and Pentland, Alex},
  booktitle = {International Conference on Machine Learning},
  year      = {2001},
  pages     = {218-225},
  url       = {https://mlanthology.org/icml/2001/ivanov2001icml-expectation/}
}