Design Principles of Learning and Inference Models with Optimal Latent Distributions

2018 Fiscal Year Final Research Report

• Project/Area Number: 15K16050
• Research Category: Grant-in-Aid for Young Scientists (B)
• Allocation Type: Multi-year Fund
• Research Field: Intelligent informatics
• Research Institution: Toyohashi University of Technology
• Principal Investigator: Watanabe Kazuho 豊橋技術科学大学, 工学(系)研究科(研究院), 講師 (10506744)
• Project Period (FY): 2015-04-01 – 2019-03-31
• Keywords: レート歪み関数 / 再構成分布 / イプシロン不感応損失 / ディリクレ過程平均法 / ミニマックス予測

Outline of Final Research Achievements

We evaluated the rate-distortion functions of practical loss functions such as epsilon-insensitive distortion measures and those defined with kernel functions, which demonstrate the performance of optimal lossy data compression systems under these distortion measures. We interpreted a recent clustering method to extract latent structure of data and the design of the prior distribution in Bayesian inference through the optimization of latent distributions, and provided their extensions and approximations. We developed a modified latent variable model for an interactive data analysis method based on data visualization, and examined its performance through its applications to some real datasets.

Free Research Field

統計的学習理論

Academic Significance and Societal Importance of the Research Achievements

レート歪み関数は歪み有りデータ圧縮の限界を示すため、イプシロン不感応損失やカーネル関数を用いた損失などの有用性が知られている歪み尺度において、圧縮法の性能評価や改良を与える際の基準が得られた。また、クラスタリング手法やベイズ推論における事前分布の設計において得られた拡張や近似法により、既存手法の個々の問題におけるより柔軟な適用や効率的な計算が可能になった。