Dr. Kaushik Chowdhury and PhD student Suyash Pradhan joined InterDigital at the 2024 Brooklyn 6G Summit to showcase their collaborative research on Over-the-Air Federated Learning (OTA-FL).
The Brooklyn Summit is a three-day wireless communications conference focused on the future of wireless technology. This years’ theme, “6G – From Vision to Action,” explores “the transition 6G is undergoing from the research to standardization stage.”
Federated Learning (FL) is a decentralized machine learning approach where devices collaborate on training models without sharing raw data. In traditional FL, devices send model updates to a central server individually, where they are aggregated to update the global model. However, because the radio spectrum—the range of frequencies used for wireless communication—is limited, sending large volumes of updates from numerous devices can be costly in terms of bandwidth.
To address this challenge, UT and InterDigital jointly demonstrated a hardware setup to show how model updates can be combined "over the air" using the principle of signal superposition. By leveraging this principle, OTA-FL allows multiple devices to transmit their updates simultaneously, combining them during transmission. This approach maximizes the use of available radio resources, reducing the need for separate transmissions and significantly conserving bandwidth.
“Instead of independently sending model updates from hundreds of devices to a central server for averaging, you're combining them during travel using common time-frequency resources. This approach helps greatly reduce spectrum usage and energy consumption at the nodes,” Pradhan explains
Chowdhury adds, “what we are trying to do is send relevant control information in ways that are highly efficient, so you use very little spectrum. In fact, you use the same amount of spectrum whether one device or 100 devices are transmitting. By overlapping the signals on each other, you are saving on the spectrum.”
Partners at InterDigital contributed to Chowdhury and Pradhan’s research by helping develop the simulation framework, as well as giving guidance on precoding strategy and methods to mitigate noise in OTA-FL training cycles.
Following the successful demonstration at the Summit, I sat down with Chowdhury and Pradhan to get their perspective on how the technology was received and what’s next for OTA-FL:
Q: How were you received at the conference? Was there any helpful feedback or surprising reactions?
PRADHAN: The concept was highly appreciated, and there was a lot of enthusiasm from attendees eager to learn more. I had visitors from top telecommunications equipment vendors and operators stop by my booth, leading to some engaging discussions. Our quantitative analysis of spectrum and energy savings resonated particularly with industry professionals, as they are actively exploring ways to improve load balancing and energy efficiency in the upcoming 6G standards. Some technical experts also showed interest in how their work on model compression and security would evolve in the context of this new paradigm. We are looking forward to collaborating more on such cutting-edge areas.
Q: What are the most important features or components of the demonstrated technology?
PRADHAN: Combining signals in a practical wireless environment brings in a couple of challenges due to propagation losses, hardware impairments and maintaining highly-precise synchronization across the multiple participating nodes. The key components of our system include:
- A precoding technique used by the mobile devices to proactively mitigate the losses and noise introduced by the wireless channel before signal combination.
- Strategic hyperparameter tuning and update computation to gradually mitigate the inherent noise
- A clock distribution system that delivers a unified time and frequency reference to all participants, ensuring precise synchronization at the symbol duration level, as defined by 5G standards.
Q: Who is the intended user for this technology? Can you provide examples of how this tech could be applied in real-world scenarios?
PRADHAN: The technology is designed for any device connected to cellular networks aiming to perform machine learning tasks in a collaborative, privacy-preserving manner. There are a plethora of use-cases for OTA-FL including object classification in autonomous vehicles, AR/VR applications for personalized experiences, and networked robotics for efficient factory automation. Additionally, we have explored FL for channel estimation to enhance wireless receiver performance, which is a relevant use-case for telecom operators.
Q: Are there any plans to further develop this demonstration?
CHOWDHURY: Our demonstrations so far involved a simplified network architecture with one base station and several associated mobile devices. We want to take this further in more complex environments with multiple base stations interfering with each other while this is happening. In addition, we want to look at how we can scale this up to multiuser multi antenna systems.
Chowdhury and Pradhan’s work on OTA-FL represents a promising solution to the communication challenges faced by 6-G networks. With support from InterDigital, the team plans to scale their demonstration for real-world applications, laying a foundation for more secure and energy-efficient wireless systems.