Recurrent Transfer Learning by Neural Network Regression for Human Balance Sensor Calibration

Abstract

This paper presents a novel human balance ability assessment device, called balance sensor, which uses the principle of Frustrated Total Internal Reflection (FTIR) to measure pressure distribution under human feet. Experiment is conducted to quantitatively calibrate the relationship between pressure and pixel value and physical model is established for this relationship. A special neural network is designed. The established physical model can be transferred into such neural network as initialization. Moreover, we invented a training paradigm called Recurrent Transfer Learning (RTL), which is especially suitable for regression by neural network. The regression performance outperforms state-of-art neural network initialization methods.