a practical designing theory for deep learning models

2017 Fiscal Year Final Research Report

• Project/Area Number: 15K00330
• Research Category: Grant-in-Aid for Scientific Research (C)
• Allocation Type: Multi-year Fund
• Section: 一般
• Research Field: Soft computing
• Research Institution: Yamagata University
• Principal Investigator: Yasuda Muneki 山形大学, 大学院理工学研究科, 准教授 (20532774)
• Co-Investigator(Renkei-kenkyūsha): Kataoka Shun 小樽商科大学, 商学部, 准教授 (50737278)
• Project Period (FY): 2015-04-01 – 2018-03-31
• Keywords: 深層学習 / 統計的機械学習 / 確率的情報処理 / 情報統計力学 / 理論解析 / アルゴリズム開発

Outline of Final Research Achievements

The aim of this research is to find a mathematical and information theoretical background of deep learning systems and to obtain effective algorithms for them in terms of the techniques in the probabilistic information processing and in the information statistical mechanics. We obtained the following results during this research period. (1) clarifying a mathematical background of the pre-training in deep Boltzmann machines, (2) a novel general theorem for restricted Boltzmann machines (RBMs) that states that, when mean-field methods are employed, inference results obtained from marginalized models are more accurate than those obtained from original models, and (3) a very fast test method for noise robustness of deep neural networks.

Free Research Field

統計的機械学習