2017 年 32 巻 4 号 p. A-GB5_1-12
This paper proposes a novel neural networks based model for learning user-independent features. In activity recognition using wearable sensors, user-independence of features could provide better user-generalization performance, enhance privacy protection, and both are important for using activity recognition techniques in a real-world scenario. However, designing such features is not an easy task, because it is not clear what kind of features become user-independent, and moreover, poor design of user-independence harms activity recognition performance.Hear, we propose User-Adversarial Neural Networks for automatically learning user-independent features. The proposed model considers an adversarial-user classifier in addition to a regular activity classifier in the training phase, and learn the features that help to distinguish the activities but obstruct to distinguish the users. In other words, the model explicitly penalizes representations for becoming user-dependent, while keeping activity recognition performance as much as possible. Our main result is an empirical validation on three activity recognition tasks regarding wearable sensor based activity recognition. The result shows the proposed model improves independence of features comparing with the regular deep convolutional neural networks in both qualitatively and quantitively. We also summarize future work for better user-generalization and privacy protection from the perspective of the representation learning.
