Deep Emergent Communication for the IoT

Abstract

Learning emergent communication remains a longstanding challenge in distributed Internet of Things (IoT) settings. The need to overcome tedious, complex design of hand-engineered communication protocols coupled with superior prediction and classification capabilities, make Deep Networks attractive for distributed, cooperative IoT settings. In such settings, sensing devices must sense, communicate and provide actuation whilst executing a resource-aware operation. Reliance on the Cloud for knowledge discovery is fraught with latency, connectivity, and bandwidth issues. We continue to see the emergence of edge-centric paradigms in which sensing devices at the network edge are endowed with intelligence. In turn, these devices are equipped with self-organization capabilities, robust real-time capabilities, reduced bandwidth requirements and greater context awareness. In this paper, we propose a novel, scalable communicating Convolutional Recurrent Neural Network (C-RNN) architecture for distributed IoT settings. Our framework automatically learns emergent communication in a purely data-driven way. Extensive experimental evaluation shows that our framework can learn to solve distributed image classification tasks, optimises for communication cost, is robust to lossy-links and can scale to multiple nodes.