User adaptation of convolutional neural network for human activity recognition

Abstract

Recently, monitoring human activities using smart-phone sensors, such as accelerometers, magnetometers, and gyro-scopes, has been proved effective to improve productivity in daily work. Since human activities differ largely among individuals, it is important to adapt their model to each individual with a small amount of his/her data. In this paper, we propose a user adaptation method using Learning Hidden Unit Contributions (LHUC) for Convolutional Neural Networks (CNN). It inserts a special layer with a small number of free parameters between each of two CNN layers and estimates the free parameters using a small amount of data. We collected smartphone data of 43 hours from 9 users and utilized them to evaluate our method. It improved the recognition performance by 3.0% from a user-independent model on average. The largest improvement among users was 13.6%.