Toward Understanding Acceleration-based Activity Recognition Neural Networks with Activation Maximization

Abstract

Although deep learning-based activity recognition using wearable sensors has been actively studied to implement smart applications such as supporting elderly care, healthcare, and home automation, techniques for understanding the inside of activity recognition networks have not yet been investigated thoroughly. In the computer vision research field, activation maximization (AM) was proposed to visualize the internal functions of networks. However, when conventional AM techniques, which are tailored to image-based recognition, are directly applied to acceleration-based activity recognition networks, meaningless and noisy signals are generated because of the difficulties in regularizing AM by directly using the values of the acceleration signals that are generated. This study proposes novel regularization techniques for AM using activity recognition networks that leverage activation values used in AM to indirectly control the acceleration signals that are generated. We evaluated the proposed method quantitatively using publicly available datasets and confirmed the effectiveness of the proposed techniques.