Approximation methods of accuracy of latent-variable estimation based on asymptotic error

2014 Fiscal Year Final Research Report

• Project/Area Number: 24700139
• Research Category: Grant-in-Aid for Young Scientists (B)
• Allocation Type: Multi-year Fund
• Research Field: Intelligent informatics
• Research Institution: Tokyo Institute of Technology
• Principal Investigator: YAMAZAKI Keisuke 東京工業大学, 総合理工学研究科(研究院), 助教 (60376936)
• Project Period (FY): 2012-04-01 – 2015-03-31
• Keywords: 数理統計学 / 最尤推定 / ベイズ推定 / 潜在変数推定 / 教師なし学習 / 半教師あり学習

Outline of Final Research Achievements

In many information science fields, one of the important tasks is to investigate the structure of data. Hierarchical statistical models, which is often used for data analysis, consists of two types of variables: observable and latent variables. The present study focused on estimation of the latent variable and its accuracy. Since the true latent variables are not available or observable, a method to evaluate the accuracy has not thoroughly been studied. We proposed a new method to calculate the accuracy from the given observable data and to design the optimal structure of the model. The result shows that the method enables us approximate the accuracy when the latent variable is well-specified. When the variable has redundancy, the method requires additional information of the variable. Then, we theoretically revealed the asymptotic behavior of the accuracy for an extension of the proposed method when some parts of latent variables are observable.

Free Research Field

統計的機械学習