Leveraging Ultra-Low-Power Wearables Using Distributed Neural Networks

Traditional deep learning models incur high computational overhead (approximately 1-30W) and are unsuitable for Ultra-Low-Power (ULP) wearable systems like smart glasses. In contrast, TinyML architectures are power-efficient but less accurate. We propose distributing a neural network (NN) between a ULP wearable node and a resource-rich hub to maintain high accuracy and low power consumption for applications like human-machine vision. Output features from the node are transmitted to the hub via low-power communication, where the remaining network runs on traditional GPUs. We introduce a Figure of Merit (FoM) to determine the optimal NN distribution point and a customized multiplier unit for ULP operation. Achieving ULP of 284μW and 9mW for different Autoencoder (AE) networks, our approach is 700× lower in power consumption compared to traditional GPU implementation.