Event
Ph.D. Dissertation Defense: Cemil Nureddin Vahapoglu
Monday, December 1, 2025
11:30 a.m.
AVW 1146
Sarah Pham
301 473 2449
spham124@umd.edu
First, we introduce an unsupervised neural beamforming (NNBF) framework for uplink multi-user single input multiple output (MU-SIMO) systems under 3GPP-compliant channel models. So, in contrast to supervised DL-based beamforming methods, we remove the reliance on ground-truth beamformers by directly optimizing the sum-rate objective in an end-to-end fashion. We propose two architectures: a lightweight convolutional design, simple NNBF, for stationary user equipments (UEs) with perfect CSI, and a transformer-based design that integrates grouped convolutions and multi-channel attention, tailored for dense urban environments with imperfect CSI and user mobility. Both models are trained by using frequency-domain channel responses, and they generate beamforming weights directly in the frequency domain. Our results demonstrate that simple NNBF achieves comparable or better performance than MMSE while consistently outperforming ZFBF across all SNR regimes with reduced complexity. Moreover, transformer-based NNBF provides further gains compared to baseline techniques ZFBF and MMSE with imperfect CSI under high-mobility conditions at the cost of higher computational overhead.
Next, we extend transformer-based NNBF by introducing Doppler-aware sparse attention, leading to the Doppler-aware Sparse Neural Beamforming (Sparse NNBF) framework. In this design, we directly tackle the quadratic complexity of standard full attention by using a Doppler-aware sparse attention mechanism, which improves the scalability for large OFDM grids and the robustness for high mobility scenarios. Specifically, we design a channel-adaptive sparse attention mechanism that changes sparsity patterns along the two-dimensional time-frequency grid according to channel dynamics, such as the Doppler shift effect. We provide a theoretical guarantee of full connectivity; any two query-key pairs can be connected within a bounded number of hops \textit{p}, where \textit{p} denotes the number of heads. This ensures that our sparsification design does not compromise expressiveness. Compared to fixed sparsity schemes such as strided or local windowing, Doppler-aware sparse adapts to temporal variability and captures wireless channel dynamics better. Our experimental findings with 3GPP UMa channels demonstrate that Sparse NNBF surpasses both traditional beamforming methods and transformer-based NNBF utilizing static sparse patterns, revealing resilience in high-mobility scenarios.
Then, we introduce NNBF-P, an unsupervised DL framework for the joint design of downlink multi-user multiple input single output (MU-MISO) beamforming and power allocation. NNBF-P learns both beamforming vectors and power allocation coefficients simultaneously by directly maximizing the sum-rate objective in an end-to-end unsupervised manner, thereby enabling a unified design that adapts to channel conditions within a single process. Our simulation results with 3GPP channel models, TDL-A and TDL-C, demonstrate that NNBF consistently outperforms ZFBF, MMSE, and NNBF without power allocation regarding spectral efficiency under varying channel delay spreads and modulation schemes.
Finally, we propose Hierarchical Over-the-Air FedGradNorm (HOTA-FedGradNorm), a distributed training framework designed for scalable and resource-efficient learning across heterogeneous wireless tasks. HOTA-FedGradNorm integrates hierarchical federated learning (HFL) with over-the-air aggregation (OTA) and dynamic task weighting. In contrast to the beamforming-centric discussions of previous parts, we tackle issues of multi-task radio learning in federated settings, where client devices and network nodes may differ in computational capabilities, channel conditions, and task complexity. We demonstrate that HOTA-FedGradNorm can achieve efficient convergence under wireless channel constraints through the dynamic balancing of task gradients and the utilization of OTA. While we do not present it as a beamforming algorithm, we describe it as a general-purpose training substrate for physical (PHY)-layer tasks, including beamforming, channel prediction, and resource allocation in O-RAN deployments. By establishing HOTA-FedGradNorm as a standalone contribution, we emphasize its function as an inspirational bridge toward deployment and feasibility, promising that neural beamforming and other data-driven PHY-layer methods can be trained collaboratively and efficiently in practice.
