Task-oriented communications and networking for learning, optimization, and human-machine symbiosis in 6G

Abstract

As the evolution towards the sixth generation (6G) of wireless networks unfolds, it is prompting the need for innovative designs and techniques to facilitate efficient, task-oriented communications. This is essential to meet the increasingly complex demands of emerging applications. Traditional wireless networks focused primarily on high-rate data transmission, however, the diverse requirements of novel applications have given rise to multi-functional networks. This dissertation presents solutions for task-oriented communications and networking specifically tailored for learning, optimization, and human-machine symbiosis in the context of 6G.

One such application is the deployment of federated learning in wireless networks, known as federated edge learning (FEEL). FEEL leverages low-latency access to spatially distributed mobile data, thereby efficiently training AI models whilst preserving data privacy. When deployed in a large-scale cellular network, the spatial learning performance of FEEL comes into focus. First part of this dissertation evaluates digital and analog transmission schemes and their distinct effects on the convergence rate in the network, particularly under dense device conditions. It identifies the constraints posed by digital transmission due to a limited number of active devices and how the spatial convergence rate for un- coded analog over-the-air computation (AirComp) can be affected by inter-cell interference. Despite the challenges, the potential of transmission using AirComp in reducing learning latency is highlighted, due to its capacity for simultaneous access.

The second segment of this dissertation focuses on distributed optimization, a technique extensively used in applications ranging from machine learning to vehicle platooning. This process often faces a bottleneck due to the extensive exchange of messages required. To address this, an innovative framework of distributed AirComp is proposed by superimposing multiple conventional AirComp processes simultaneously. This necessitates effective multicast beamforming design at devices to control errors stemming from interference and channel distortion. Two design criteria are proposed and investigated: minimizing the sum mean-squared error (MSE) with respect to the desired average-functional values, and the zero-forcing (ZF) multicast beamforming. Consequently, this approach realizes a one-step aggregation for distributed optimization, dramatically accelerating convergence by reducing communication latency.

In the context of human-machine symbiosis, the 6G paradigm underlines the importance of semantic communications, which employs computation-communication integrated designs to augment semantic accuracy and task effectiveness. This methodology is vital in human-to-machine communications, centralizing efficient and effective semantic information transfer. However, achieving accurate encoding of complex semantic information into machine-readable format can trigger communication bottlenecks. To address this, the third part of this dissertation presents a novel quantization method called semantic importance-aware bit allocation (SIMBA), a quantization method to lessen semantic loss during message encoding. SIMBA decomposes word embedding vectors into blocks of varied semantic importance, thus informing a unique bit allocation strategy. Additionally, a task-specific adaptive quantization scheme is introduced. This scheme optimizes codebook size based on command type frequencies to minimize message size and reduce communication overhead. The implementation of the proposed schemes effectively preserves semantic information and reduces communication overhead, achieving a balance between semantic accuracy and efficiency in ultra-low latency transmissions.