2018 Fiscal Year Final Research Report

New framework of multiple kernel learning

Research Project

PDF

Project/Area Number	16K05264
Research Category	Grant-in-Aid for Scientific Research (C)
Allocation Type	Multi-year Fund
Section	一般
Research Field	Foundations of mathematics/Applied mathematics
Research Institution	Hokkaido University
Principal Investigator	Tanaka Akira 北海道大学, 情報科学研究科, 教授 (20332471)
Project Period (FY)	2016-04-01 – 2019-03-31
Keywords	機械学習 / カーネル回帰 / 再生核 / 再生核ヒルベルト空間 / 汎化性能
Outline of Final Research Achievements	We obtained the following two results in kernel-based machine learning. 1. We analyzed the theoretical limit of conventional multiple kernel learning, and clarified that (a) the optimal solution based on an ordinary 2-norm criterion can not achieve the theoretical limit of the multiple kernel model, and (b) the optimal solution is identical to the solution based on a certain single kernel model. 2. We developed a novel framework of kernel-based learning in which the autocorrelation prior of the unknown true function is taken into account, and also clarified that the autocorrelation function itself is the best kernel in terms of the expectation.
Free Research Field	機械学習
Academic Significance and Societal Importance of the Research Achievements	現代社会において、機械学習理論が果たす役割がますます増大するであろうことは最早論を俟たない。一方、それら技術の実用面での爆発的な発展とは裏腹に、理論的な背景や基盤が置き去りにされていることは、当該分野の未来にとって由々しき事態であると言わざるを得ない。本研究課題で得た成果は、主に、機械学習理論における理論整備として位置づけられるものであり、当該分野の科学的な側面を補強する一助になるものと期待できる。