Artificial Intelligence in the Air

Intelligence is coming to the edge. Autonomous systems and industrial edge are the next targets of artificial intelligence (AI). However, despite impressive progress in hardware, edge devices do not have sufficient computing power and data to train and deploy state-of-the-art machine learning (ML) algorithms. Communication can allow edge devices to share their data and computational resources, and provide manifold increase in their learning capabilities, similarly to the impact language had on human intelligence. However, influenced by Shannon's seminal work, our current communication architectures are designed to establish reliable bit pipes between nodes, dismissing the relevance or utility of delivered bits. Yet, in ML applications, we are interested in inferring features of the underlying signals or messages, rather than reconstructing them. Increased data rates do not translate into faster or more accurate learning algorithms, and the content, timeliness, and the relevance of information are often more important than its quantity. AI-R challenges the current framework that treats communication and learning separately, striving to bridge this gap by developing an AI-oriented communication paradigm from fundamental theoretical principles. This new paradigm will go beyond the classical communication-theoretic framework by taking into account the ultimate goals of information transmission, for example, detecting anomalies in drone footage or remote controlling an industrial robot. AI-R will also step out of the cycles of incremental research in communications by fully exploiting AI capabilities to `learn' the best communication strategies to achieve the prescribed objectives. Building upon our expertise and recent contributions in information theory, coding, communications and ML, the project will balance fundamental research with application-oriented algorithm design and implementation to develop new engineering insights and products towards ambient edge intelligence.