Rule Generation from Wrist EMG Recognition Network Using Deep Learning and Muscle Synergy to Increase Data Value

2022 Fiscal Year Final Research Report

• Project/Area Number: 20K12600
• Research Category: Grant-in-Aid for Scientific Research (C)
• Allocation Type: Multi-year Fund
• Section: 一般
• Review Section: Basic Section 90110:Biomedical engineering-related
• 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): 2020-04-01 – 2023-03-31
• Keywords: 手首筋電 / データ増量 / 深層学習 / 筋シナジー

Outline of Final Research Achievements

In the input layer, the training data is multiplied by Gaussian random numbers, the number of data wass increased several hundred times, and their values (usefulness) were evaluated in an abstracted space near the middle layer. After learning, dimensionality reduction was performed by t-SNE near the hidden layer. It was found that learning using data below a certain threshold had the effect of improving recognition accuracy.
Next, a genetic algorithm was used to determine the frequency distribution of myoelectric signals for each muscle synergy. As a result, some differences could be detected for each muscle synergy.
Finally, regarding rule extraction, we have not been able to develop an effective extraction method for convolutional networks. Currently, we are studying a rule extraction method by connecting three-layer neural networks.

Free Research Field

信号処理

Academic Significance and Societal Importance of the Research Achievements

入力層で学習データ数を増加させ、その価値（有用度）を中間層付近の抽象化された空間で評価する方法は、従来とは異なる新しい手法であり、またその効果も大きい。この方法は深層学習を用いるすべての分野に適用することが可能で有効性と影響度は大きい。今後、様々な分野に適用していく予定である。
深層学習ネットからのルール抽出法の開発は、深層学習ネットの内部のホワイトボックス化に繋がり、大変意義の大きい研究である。現時点では、有効な解決法を見いだせていないが、今後、3層ネットの連結化などの方法を検討し、ルール抽出の研究開発を行う予定である。