ゼロショット行動認識のための中間表現の探索

Masashi Takeshita, Moe Matsuki, Sozo Inoue,
日本知能情報ファジィ学会九州支部学術講演会予稿集
(Not Available)
(Not Available)
6-9
2019-11-30
Iizuka
http://soft-kyushu.org/2019/conf/index.html
Activity recognition technology is required in ubiquitous computing. In recent years, activity recognition technology uses machine learning, but machine learning has two problems; extracting feature of sensor data is required to carry out by hand and the data collection task is quite laborous. To overcome these problems, we aim to propose the method combining deep learning model and Zero-shot learning model. However, we are worried that by combining these approach, the performance of unknown activity class recognition become decreases. In this paper, we study how to recognize attribute vector in the middle layer for unknown activity classes recognition. At the result, proposed model which recognize single-attribute in the middle layer achieve 60 % improvement
compare to the multi-label classification neural network model.

Data Files

  • スクリーンショット_2019-11-30_11.12.36.png