Event

ANNOUNCEMENT: Ph.D. Dissertation Defense

 

Name: Matin Mortaheb

 

Committee: 

Professor Sennur Ulukus (Chair)

Professor Behtash Babadi

Professor Adrian Papamarcou

Dr. Mohammad A. Khojastepour

Professor Dinesh Manocha (Dean's representative)

 

Date/time: Monday, July 21, 2025 at 11:00 AM

 

Location: AVW 1146

 

Title: Deep Learning-Enabled Intelligent Goal-Oriented and Semantic Communication for 6G Networks

 

Abstract: 

In the emerging sixth generation (6G) wireless networks, two major technological directions are expected to play pivotal roles: semantic communication and federated learning. Although independent in their approaches, both aim to make networks more intelligent, efficient, and adaptive to user and application demands. Semantic communication fundamentally shifts the focus from transmitting raw data to conveying only information that carries meaning and is directly relevant to a given task. This shift enables more efficient use of spectrum and reduces unnecessary transmission overhead, which is critical for supporting services such as augmented reality, holographic telepresence, and real-time control systems. In parallel, federated learning enables distributed edge devices to collaboratively train powerful machine learning models without sharing their raw data, thereby ensuring privacy, reducing communication burdens, and supporting large-scale, decentralized intelligence. Together, these advancements address key 6G challenges such as ultra-low latency, massive connectivity, and the need for real-time, context-aware decision-making at the edge. By integrating semantic understanding and distributed learning capabilities into the communication fabric, 6G networks will evolve from simple data carriers into intelligent infrastructures that empower a new generation of applications and services. This thesis addresses the design, analysis, and integration of advanced techniques that enable intelligent and efficient communication across multiple emerging 6G applications, including distributed learning, real-time immersive media, and multi-modal AI systems.

 

A primary contribution of this thesis lies in the development of semantic communication frameworks, where data is encoded and transmitted in a manner that preserves semantically important content while discarding redundant or irrelevant information. We propose an adaptive semantic encoder that utilizes attention mechanisms from vision-language models to dynamically identify and prioritize semantically salient regions in images and videos. By allocating communication resources according to semantic importance, our framework significantly reduces bandwidth consumption and latency, enabling critical 6G use cases such as holographic telepresence, remote surgery, and augmented/virtual reality with enhanced reliability and efficiency.

 

Parallel to semantic encoding, this thesis investigates federated learning (FL) as a foundational pillar for distributed intelligence in 6G. In FL, multiple devices collaboratively train a shared model without sharing their raw local data, offering strong privacy guarantees. However, practical challenges arise due to severe data heterogeneity across devices and limited wireless resources. To address these, we develop FedGradNorm, a dynamic gradient normalization scheme that harmonizes learning speeds across heterogeneous tasks, ensuring fair and efficient convergence in personalized federated multi-task learning (PF-MTL) settings. Building upon this, we introduce HOTA-FedGradNorm, which leverages over-the-air aggregation (OTA) to enable simultaneous and bandwidth-efficient transmission of local model updates to a centralized parameter server. Furthermore, we propose a novel graph-based decentralized learning approach that adaptively adjusts communication topologies by automatically learning inter-client task correlations, thereby minimizing negative interference and enhancing convergence performance. These contributions enable FL frameworks that are communication-efficient, robust to statistical heterogeneity, and tailored for highly dynamic and bandwidth-limited 6G environments.

 

In addition to semantic encoding and federated learning, this thesis addresses the challenge of adaptive video streaming over wireless networks, a critical service for 6G that supports ultra-high-definition and interactive media applications. Traditional rate control algorithms lack the ability to dynamically adapt to rapidly changing network conditions and user experience requirements. We design a machine learning-based rate control algorithm that employs deep learning to jointly optimize video bitrate selection by considering real-time channel states, buffer dynamics, and semantic content complexity. Our approach achieves superior perceptual quality and stable playback, providing a seamless user experience even under highly variable wireless channel conditions.

 

Lastly, to ensure reliable and trustworthy information delivery in complex 6G multi-modal systems, we introduce RAG-Check, a novel framework for evaluating and mitigating hallucinations in multi-modal retrieval-augmented generation (RAG) systems. As 6G networks are envisioned to support advanced AI services that integrate visual, textual, and contextual data, ensuring the factual correctness and semantic alignment of generated responses becomes critical. RAG-Check rigorously analyzes the consistency between retrieved image evidence and generated textual outputs, providing quantitative and interpretable hallucination scores. This framework enhances the robustness and credibility of large-scale vision-language models in real-world deployments.

 

By cohesively integrating these contributions, this thesis offers a comprehensive design blueprint for goal-oriented and semantic communication systems in 6G networks. The thesis bridges the gaps between intelligent content encoding, collaborative distributed learning, adaptive multimedia delivery, and reliable multi-modal AI, collectively advancing the vision of 6G networks that do not merely transmit data but understand, reason, and act on information to effectively meet diverse user goals and application requirements, thereby laying the foundations for future intelligent communication systems that are efficient, scalable and also contextually aware and user-centric, enabling the next era of ubiquitous and intelligent connectivity.