AI in NextG Networks: from Neuromorphic Hardware to Applications

The increasing complexity and demand for ultra-low latency, high data rates, and massive device connectivity in Next Generation (NextG) wireless networks require new paradigms for network management. As wireless technologies evolve, the integration of Artificial Intelligence (AI) becomes essential to effectively address the dynamic nature of NextG systems. AI, particularly Deep Learning (DL), offers the potential to automate resource management, optimize performance in real time, and enable emerging applications, such as immersive augmented reality, holographic telepresence, and unmanned mobility. This thesis introduces a comprehensive multi-layered approach to integrating AI tech- nologies into NextG networks, spanning from neuromorphic photonic hardware to middle- ware and system-wide implementation. At the hardware level, we introduce Photonic-Aware Neural Networks (PANNs), a novel class of neural networks that leverage the parallelism, speed, and energy efficiency of neuromorphic photonic accelerators. PANNs demonstrate significant improvements over traditional electronic systems in high-throughput, low-power processing, essential for AI-driven tasks in NextG networks. We demonstrate their applica- tions in computer vision and network security tasks, tackle the noise challenges introduced by photonic accelerators, and explore hardware implementations. Moving to the middleware level, we discuss the integration of In-Network Machine Learning (ML) techniques that allow real-time decision-making directly within the network infrastructure. By deploying ML models in network devices such as switches and routers, the network can perform tasks like traffic management and cybersecurity threat detection with reduced latency and computational load. This approach enhances the overall intelligence and responsiveness of the network by processing data directly within the network devices. Finally, at the system-wide orchestration level, we propose a converged Radio Access Network-Core Network (RAN-CN) architecture, enabling real-time AI-driven analytics directly at the base station. In addition, we introduce AI-powered resource management techniques that dynamically scale network resources based on traffic forecasting, ensuring efficient pro-active resource utilization. Furthermore, we investigate Federated Learning (FL) techniques for distributed AI model training at the edge of the network, supported by Fountain Codes (FC) to ensure reliable model updates even under challenging network conditions. Through these layers of integration, this thesis demonstrates how AI solutions can enhance the scalability, efficiency, and sustainability of NextG wireless networks, making them more adaptive and capable of addressing the demands of future wireless applications.