Optimized deep learning architectures for radio intelligence in 6G networks

Abstract

The transition from Fifth Generation of Cellular Networks (5G) and 5G-Advanced toward Sixth-Generation Mobile Networks (6G) is redefining wireless communication systems as Artificial Intelligence (AI)-native infrastructures, where intelligence is embedded throughout the communication architecture and pushed closer to the network edge. In this context, the Physical Layer (PHY) emerges as a key domain for the integration of data-driven models capable of supporting prediction, classification, and adaptive decision-making under strict constraints on latency, memory, energy consumption, and computational capacity. However, a persistent gap remains between high-performance Deep Learning (DL) models developed in unconstrained computing environments and the requirements for their practical deployment on heterogeneous edge devices. This thesis addresses this gap by developing a unified framework that integrates efficient model design for the radio domain, complexity reduction, hardware-aware optimization, deployment, and experimental validation on computationally constrained Radio Access Networks (RAN)-oriented platforms for AI-native systems. The research is structured as a closed methodological loop that begins with the design of efficient DL architectures for radio-domain tasks and progresses through model compression, Multi-Task Learning (MTL), layer-wise quantization, and automated architecture search. The process culminates in deployment and experimental validation on embedded processors, accelerators, and programmable logic devices. The central objective is to transform computationally intensive models with limited deployability into solutions capable of operating under realistic far-edge conditions while preserving competitive predictive performance and reducing computational and deployment costs. In line with this objective, the thesis first proposes a lightweight channel-gain prediction pipeline for distributed radio access environments, combining Channel State Information (CSI) compression, compression-aware training, and attention-enhanced recurrent prediction. This approach reduces both transport-layer overhead and model computational complexity while maintaining bounded degradation in prediction accuracy and enabling Real-Time (RT) execution under practical latency constraints. Second, a systematic methodology is developed for spectrum-based Traffic Classification (TC) from In-phase and Quadrature components (I/Q) samples using compressed multi-task architectures. By exploiting shared representations across related tasks, the proposed models reduce redundancy, memory footprint, inference time, and energy consumption while preserving accuracy levels compatible with real-time edge operation. Third, the thesis introduces a methodology for layer-wise quantization of Deep Neural Networks (DNNs) for radio receivers. Through layer-level sensitivity analysis, the proposed framework determines mixed-precision configurations that better balance computational efficiency and predictive performance than uniform quantization schemes, thereby facilitating implementation on hardware platforms with limited computational capacity. Fourth, a resource-aware neural architecture search framework is presented to jointly optimize convolutional architectures and adaptive quantization under hardware-relevant constraints. The method integrates Particle Swarm Optimization (PSO), surrogate modeling, and reduced-data evaluation to efficiently explore large design spaces. The resulting architectures approach the Pareto frontier between accuracy and computational cost while substantially reducing search time and dependence on manual design decisions. Finally, the models obtained through this design loop are deployed and experimentally validated on computationally constrained devices, including accelerated Graphics Processing Unit (GPU) environments and Field-Programmable Gate Array (FPGA)-based platforms. These experiments confirm that the proposed hardware-aware strategies translate theoretical reductions in complexity into practical gains in latency, power consumption, throughput, and logical resource usage. In this way, the thesis demonstrates that computationally intensive DL solutions can be transferred from algorithmic conception to real deployment on constrained platforms and realistic wireless applications. This doctoral thesis contributes a systematic framework for bridging the gap between high-accuracy yet computationally demanding models and deployable AI solutions for edge intelligence in future wireless networks. By jointly addressing efficient model design, complexity reduction, hardware-aware optimization, deployment, and experimental validation, the thesis advances the practical integration of efficient DL techniques into AI-native RAN architectures and next-generation wireless communication systems.