Information-Theoretic View and Development of Kernel-Based Learning Methods

2014 Fiscal Year Final Research Report

• Project/Area Number: 23700175
• Research Category: Grant-in-Aid for Young Scientists (B)
• Allocation Type: Multi-year Fund
• Research Field: Intelligent informatics
• Research Institution: Toyohashi University of Technology (2014); Nara Institute of Science and Technology (2011-2013)
• 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): 2011-04-28 – 2015-03-31
• Keywords: 混合モデル / レート歪み関数 / 再構成分布 / イプシロン不感応損失 / 漸近的ミニマックス性

Outline of Final Research Achievements

We developed a learning method of mixture models, which unifies existing methods such as convex clustering and kernel vector quantization. We provided an interpretation of this method as an optimization problem for the evaluation of the rate-distortion function. The rate-distortion function indicates the performance of optimal lossy data compression systems. We evaluated rate-distortion functions of practical loss functions such as absolute-log distortion and epsilon-insensitive distortion measures. We also evaluated the performance of an online prediction algorithm to provide an efficient method to approximate the optimal prediction algorithm.

Free Research Field

統計的学習理論、機械学習