Preliminary Investigation of Visualizing Human Activity Recognition Neural Network

Abstract

Owing to the growing demand for wearable context-aware applications, activity recognition technologies have attracted great attention. A neural network has been recently used as a recognition algorithm because of its discrimination and feature extraction ability. While understanding the network provides us useful information to improve its performance, visualization techniques for neural networks have been not explored yet in the human activity recognition field. We propose a visualization method tailored to human activity recognition that generates acceleration signals which maximize the activation of a unit in a neural network. We introduce a new regularization method based on a low pass filter to suppress high-frequency components induced in the generation process to improve the interpretability of the signals.