Construction of Innovative Interface Platform by Wrist EMG based on Rule Extraction from Deep Net

2019 Fiscal Year Final Research Report

• Project/Area Number: 16K01357
• Research Category: Grant-in-Aid for Scientific Research (C)
• Allocation Type: Multi-year Fund
• Section: 一般
• Research Field: Biomedical engineering/Biomaterial science and engineering
• Research Institution: The University of Tokushima
• Principal Investigator: FUKUMI Minoru 徳島大学, 大学院社会産業理工学研究部(理工学域), 教授 (80199265)
• Project Period (FY): 2016-04-01 – 2020-03-31
• Keywords: 手首筋電 / ジャンケン認識 / 深層学習 / 個人認証

Outline of Final Research Achievements

In this study, we proposed a method for increasing the number of learning samples by using random numbers for deep learning nets that perform rock-paper-scissors recognition, and evaluated its effectiveness. In addition, we proposed a method of learning recognition by dividing the output layer part from the middle layer of the deep learning net into two networks: class classification and personal authentication (subject classification), and achieved high accuracy.
We have examined various mechanisms for rule extraction, but as a result, we have not been able to develop a mechanism that can effectively extract identification rules by deep learning with convolutional layers. In the future, we will study the development of a new mechanism.

Free Research Field

信号処理

Academic Significance and Societal Importance of the Research Achievements

本研究では，ジャンケン認識を行う深層学習ネットに対して，乱数を用いて学習用サンプル数を増加させる方法を提案し，その有効性を評価した．また，深層学習ネットの中間層付近から出力層部分をクラス分類と個人認証を行うネットワークの二つに分けて学習認識する方法を提案し，高精度を達成できた．これらの仕組みは，他の分野で使用される深層学習ネットにも直接適用可能な方法であり，また個人認証を含むことからセキュリティ面での安全性強化に役立つ方法である．