Can smart devices cooperate to train artificial intelligence (AI) models when there is insufficient internet connection? Yes, and Xiaowen Gong, Godbold Associate Professor of Electrical and Computer Engineering, can prove that.
Gong's recently completed National Science Foundation-funded study, “Quality-Aware Distributed Calculation for Wireless Federated Learning: Channel-Aware User Selection, Mini-Batch Size Adaptation, and Scheduling,” demonstrates how smart devices can build better AI models regardless of connection quality and turn network constraints into manageable constraints.
Originally funded and commissioned in 2021, his work paves the way for smarter, faster and safer technology. This encourages innovation to make robots more capable, make augmented/virtual reality experiences more immersive, and make autonomous and wireless systems more intelligent.
“Our algorithms allow federated learning in wireless networked systems where devices are unreliable, time-varying, heterogeneous communication and computational capabilities,” Gong said. “Our research will improve learning accuracy and accelerate the training process, but everything will give the device more flexibility and allow it to participate.”
Federated Learning allows multiple devices, such as smartphones, tablets, and sensors, to collaborate and train AI models without sharing raw data. Instead of sending sensitive information to a central server, the device processes the data locally and shares only learning updates. This approach protects privacy while enabling AI systems to learn from a variety of data sources.
“AI doesn't just live in large data centers,” says Gong. “It happens on devices that we use every day, such as phones, cars, smart home systems,” says Gong. “Our work helps these devices learn together, even if the internet connection is not perfect. This means smarter predictions, faster responses, and better performance in real terms.”
Existing federated learning methods often do not work well when the device has unreliable connections or different computational features, resulting in slower training and more accurate models.
Gong's work addresses this problem via a method described as quality recognition variance calculations. The new algorithm intelligently selects the devices that participate in each training round and adjusts the amount each device performs based on the quality of the connection and computational power.
“Our method not only improves the learning accuracy of federal learning, but also accelerates the training process, allowing devices to participate with a lot of flexibility, even if some devices come and go,” he said.
“Imagine a smart assistant learning new things 30% faster or a car responds quickly to changes in traffic. That's an improvement we see. This isn't just speed. It's about making AI more responsive and reliable in everyday life.”
