Research Motivation

Over-the-Air Computation (AirComp) exploits the natural waveform superposition property of wireless multiple-access channels, allowing multiple devices to transmit their local data simultaneously and enabling functions such as summation and averaging to be computed directly over the air. Unlike conventional communication systems, AirComp does not require the receiver to first decode the data from each individual device and then perform the computation. This paradigm is particularly attractive for applications such as massive IoT, distributed sensing, and federated/edge learning, where the server often does not need to recover the raw data from every device, but only requires an aggregated function of distributed data. As the number of devices grows, conventional orthogonal multiple access incurs substantial communication overhead and multi-access latency. By integrating communication and computation, AirComp can significantly reduce the latency of large-scale data aggregation and provide an efficient wireless interface for distributed intelligence.

Main Challenges

The key challenges of AirComp include the following. Wireless fading and the near–far effect can destroy the amplitude alignment required for accurate signal superposition at the receiver. Noise, interference, and imperfect channel state information (CSI) directly translate into computation errors. In addition, achieving sufficiently accurate time synchronization among a large number of distributed devices remains a critical implementation challenge. Another major issue is that classical AirComp largely relies on uncoded analog transmission, which creates a considerable implementation gap with existing communication systems based on digital modulation and channel coding. In particular, analog AirComp is inherently sensitive to noise and interference, making computation reliability a major concern for high-reliability 6G sensing, learning, and applications. Improving reliability while maintaining compatibility with practical digital communication architectures is therefore an important direction for further development.

Our Work

Up to now, we has developed a relatively complete research trajectory in AirComp. Our early work focused on the physical-layer foundations of AirComp, including MIMO AirComp, clustered MIMO AirComp, multi-cell power control, and distributed optimization, with the goal of enabling efficient and reliable over-the-air function computation through beamforming, power control, and distributed algorithm design. We subsequently introduced AirComp into federated edge learning (FEEL) and developed broadband analog aggregation and one-bit over-the-air aggregation, allowing a large number of edge devices to directly aggregate gradients or model updates over a shared wireless channel. These works characterized the benefits of AirComp from both communication-latency and learning-convergence perspectives. Building on these foundations, we further developed digital AirComp to address the limited compatibility of conventional analog AirComp with practical digital communication systems and its vulnerability to computation errors. More recently, we have extended AirComp toward distributed sensing, in-network AI, and integrated sensing and edge AI, exploring the deeper integration of wireless communication, computation, sensing, and AI at the network edge.

Highlights

Prof. Huang’s group pioneered the Over-the-Air Computing (AirComp) technology that leverages the channel waveform superposition property to transform the air interface into computer. They published numerous highly cited papers and is among the first to apply AirComp in federated learning systems, creating the hot area of “over-the-air federated learning.” His work entitled “Broadband analog aggregation for low-latency federated edge learning” has been cited 900+ times. In recognition of his pioneering work on AirComp, he has been invited by Nature Reviews Electrical Engineering to contribute an article sharing his perspective on its practical impact on communication and computing.