Toward Projection Learning between Sensor Data and Semantic Word Vector for Zero-shot Learning

摘要

In this paper, we compare 4 learning projection models between sensor domain and text domain for Zero-shot learning (ZSL). In traditional activity recognition with sensor data, the task of collecting training dataset is too tough and costly to apply for social. Our challenge is making the task efficient. The Zero-shot learning's purpose is to recognize the unknown activity which is activity class out of training dataset. In our previous research, we propose the Zero-shot learning method using the word vectors made from Wikipedia corpus for recognizing the human living activities like breakfast, watching TV, etc. We found that this method success to recognize unknown activities and need to improve the projection function for performance. In this paper, we construct 4 learning models for projection and evaluate them with accelerometer sensor data annotated simple activities. As a result, we realize that (1) the learning method with twice projection is useful for performance. (2) It is difficult to identify the two unknown activities whose distance from known activities is closer than that between other combination of two unknown activities and the known ones.