MicroDeep: In-network Deep Learning by Micro-Sensor Coordination for Pervasive Computing

The conventional wireless sensor networks are designed to gather information from sensors, and pervasive computing systems involve such wireless sensor networks that consist of a number of tiny-/micro-sensors to sense the real-world objects and phenomena. In addition, they also need to conduct training tasks for highly-intelligent processing like pattern recognition and anomaly detection, but doing such tasks in cloud servers may often be impossible or expensive due to lack of communication facility, data transfer overhead, privacy concerns and so on. In this paper, we propose a unique approach to carry out deep learning (both training/testing) among wireless sensors nodes for in-situ, edge-heavy processing. The idea is to appropriately assign neurons of CNN (Convolutional Neural Network) to wireless nodes, each of which has limited processing capability but can have some power when they are united. For this purpose, we have designed a distributed, restricted version of CNNs. We conducted two experiments using real data; one is for anomaly detection of temperature in an over-1,400m
²
lounge space using 50 temperature sensors to confirm the learning capability as well as communication overhead, and another is for activity recognition using a 6x6 array of thin-, energy-efficient film-type infra-red sensors with micro-processors to demonstrate our concept.