Effect of Regularization-based Continual Learning in Instance-incremental Learning Scenario on Long-term Surface Electromyogram-based Pattern Recognition

Yuto OKAWA, Suguru KANOGA, Takayuki HOSHINO, Shin-nosuke ISHIKAWA
Vol. 13 (2024) p. 363-373

The development of data measurement and pattern recognition modules using wearable sensing technology and deep-learning models has encouraged the realization of daily and long-term usable surface electromyogram (sEMG)- based human-machine interfaces. Continual learning methods enable pretrained models to learn new information incrementally without forgetting previous knowledge Instance-incremental learning (instance-IL) is the scenario in which the distribution and quality of data change slightly over multiple days The effect of continual learning in instance-IL on sEMG-based pattern recognition is unclear Thus, this study investigated the effect of continual learning in instance-IL on long-term sEMG-based pattern recognition Three regularization-based continual learning methods; namely, elastic weight consolidation, synaptic intelligence (SI), and learning without forgetting, were employed to update the parameters of the pretrained/backbone model day by day In addition, a convolutional neural network-based backbone model was employed to evaluate the effect of the continual learning methods Based on an evaluation including whether the past is forgotten, continual learning improved the performance as opposed to continual fine-tuning Furthermore, SI had the best averaged performance among the three regularization-based continual learning methods These results demonstrate that regularization-based continual learning could be effective in realizing daily and long-term usable sEMG-based human-machine interfaces.

留言 (0)

沒有登入
gif