Theoretical analysis of cross-validation(CV)-EM algorithm

Authors:
Takashi Takenouchi;Kazushi Ikeda
Affiliations:
NAra Institute of Science and Technology, Ikoma, Nara, Japan;NAra Institute of Science and Technology, Ikoma, Nara, Japan
Venue:
ICANN'10 Proceedings of the 20th international conference on Artificial neural networks: Part III
Year:
2010

Citing 2
Cited 0

Pattern Recognition and Machine Learning (Information Science and Statistics)

Pattern Recognition and Machine Learning (Information Science and Statistics)
Cross-validation and aggregated EM training for robust parameter estimation

Computer Speech and Language

Quantified Score

Hi-index	0.00

Visualization

Abstract

Expectation-Maximization (EM) algorithm is a typical method to estimate parameters of a model with hidden variables and is widely used for many applications. The EM algorithm is simple but sometimes overfits to specific examples and its likelihood diverges to infinite. To overcome the problem of overfitting, Shinozaki and Osterndorf have proposed the CV-EM algorithm in which the cross-validation technique is incorporated into the conventional EM algorithm, and have demonstrated validity of the algorithm with numerical experiments. In this article, we theoretically investigate properties of the CV-EM algorithm with an asymptotic analysis and reveal its mechanism of robustness.