Distributed Learning and Inference Systems: A Networking Perspective

Artificial intelligence (AI) has made significant strides, achieving and in some cases surpassing human-level performance. This has primarily been accomplished through the centralized training of static models that are then stored in centralized clouds for inference. Centralized approaches present several challenges, including privacy concerns, high storage demands, vulnerability to single points of failure, and substantial resource requirements. These limitations sparked interest in developing decentralized approaches to alleviate some of these shortcomings. Yet, decentralization introduces additional complexities, particularly in managing multiple dynamic components. Regardless of whether AI systems are centralized or decentralized, it is clear that a robust enabling infrastructure is essential for reliable and scalable operation. While simpler infrastructures may suffice for centralized approaches, distributed learning and inference require more sophisticated architectural designs. To address this gap, this paper proposes a network-inspired distributed AI service architecture, termed as Data and Dynamics-Aware Inference and Training Network (DA-ITN), designed to support mobility and decision-making across diverse AI scenarios. The components and functions of DA-ITN are explored, its potential role in the future of AI is discussed, and the various challenges and research opportunities required to realize such an architecture are identified.