Ultra-low Power Analog Recurrent Neural Network Design Approximation for Wireless Health Monitoring

Recently, the trend of analyzing physiological mark-ers for health tracking using wearable sensors is on the rise. However, due to the small size of these wearables, battery-life is of paramount concern both because of user-experience and the continuity of monitoring. Unlike the heavy mobile devices, which can be packed with powerful batteries, wearable sensors cannot, therefore, in this paper we present an ultra-low power analog design for physiological signal processing showing the potential of operating without battery or just by storing energy in the capacitors. In this work, a brand-new concept of an all-analog Recurrent Neural Network (RNN) is presented. An analog oscillator is designed to serving as the timing signal to trigger and sequence the Resistive Processing Units (RPUs) crossbar array and analog memory in a feedback loop. We evaluate the performance via an ECGs and breathing database labeled with diseases. The results show the analog RNN can classify the disease in a training accuracy of 95.57% and test accuracy of 94.16%. The architecture of our RNN consists of 200 LSTM cells with an embedding dimension of 500.