NSF Grant Trifecta Accelerates Machine Learning in Wireless Communications

From mobile health monitoring to online learning and remote work, the COVID-19 crisis has heightened reliance on wireless networks. This “new normal” needs and feeds emerging technologies and applications. Well before the pandemic, U.S. government agencies began to invest in long-term research and development in wireless communication to meet national priorities. Amid the pandemic, transferring early-stage research to the private sector is vital. “We are seeing a trend toward both fundamental and applied research inspired by real-world problems with a seamless transition to industry,”Cong Shen, University of Virginia assistant professor of electrical and computer engineering, said. Shen is an expert in machine learning for wireless networks and possesses extensive industry R&D experience and 17 U.S. patents in wireless communications and networking. He translates theoretical models into practical engineering applications, clearing the path toward more intelligent wireless communications. With partners at Penn State and the University of Miami, Shen has earned three grants from the National Science Foundation in 2020 to meet rising demands on wireless networks and advance machine learning. The cumulative investment of the awards is approximately $1.2 million, of which nearly $700,000 is allocated to Shen's research at UVA. Intel's partnership with the National Science Foundation is one indication of industry's appetite for new methods such as machine learning. The partnership program awarded a grant to Shen and Jing Yang, assistant professor of electrical engineering and computer science at Penn State, to help wireless network operators and service providers reliably meet user demands for virtual reality and high-res video applications, embedded and wearable tech and large-scale infrastructure for smart homes and autonomous cars. Shen and Yang combine domain knowledge of wireless networking with algorithms that learn via interaction and feedback to enhance network reliability and simplify its management. Their research also accommodates the exponential growth of wireless networks. Knowledge gained in one network can be easily transferred to other networks that vary widely in customer use and the “7Vs” of big data: volume, velocity, variety, variability, veracity, visualization and value. “Knowledge transfer and efficient exploration are really the key components of our proposed solution,” Yang said. “Humans do not start a new task from scratch, and we want wireless networks to behave this way, too.” The recent industry trend to leverage the computation power of smartphones, tablets and “internet of things” devices is another driver for applied research. Google and other industry leaders are exploring an emerging machine learning paradigm called federated learning. Federated learning trains a machine learning algorithm across multiple decentralized devices without uploading any data to the server. This machine learning technique helps the algorithm become more savvy while satisfying privacy and security constraints. Federated learning relies on the existing communication and networking infrastructure. Shen is partnering with Jie Xu, assistant professor at the University of Miami, to design a novel communication system that supports the unique characteristics of federated learning not captured by Wi-Fi or LTE. A grant from the National Science Foundation's communications, circuits and sensing-systems program supports this research. “We address a potential blind spot of machine learning researchers who assume that wireless communication occurs in a near-perfect, steady state,” Xu said. “In reality we know this is not the case.” Current wireless protocols are designed with instantaneous voice and data delivery in mind. Federated learning, by contrast, deploys machine learning models to devices at the edge of the network over a certain period of time. “This critical difference leads us to re-think how communication is designed,” Shen said. “Federated learning only cares about communication of machine learning models. They are very different than the typical voice or data that flows through a Wi-Fi, LTE, or 5G network, which makes the de facto choice highly suboptimal and also opens the door for innovation.” The ecosystem for wireless communication is getting bigger, denser and wilder. Network users compete for wireless spectrum, the ecosystem's most precious resource. At the same time, the operations of non-commercial services such as astronomy, atmospheric and geospatial science, weather radar and the Global Positioning System should be well protected. Shen and Yang have joined forces to study dynamic spectrum access with funding from NSF's program for spectrum and wireless innovation enabled by future technologies, known as SWIFT. Shen and Yang seek to create a novel online learning-based framework for low-cost electronic devices to efficiently and effectively access shared spectrum. “To this day, solutions to spectrum allocation rely on sensing. To avoid interference, a device must listen in on the band of spectrum it wants to use before it begins to speak,” Shen explained. However, many of the low-cost devices that compose the internet of things lack a powerful receiver RF front end, rendering them deaf to others' wideband signals. Shen and Yang will customize a machine learning tool called multi-armed bandits to help these devices find open spectrum. It will enable wireless devices to learn from the success or failure of past communications to estimate the value of a particular channel at a particular time. “Multi-armed bandits has found many successes in other fields, such as recommender systems and clinical trials,” Yang said. “We feel that this is another perfect application for it to play a significant role in solving the challenging problem of no-sensing spectrum access.” Shen and his partners focus on a few key building blocks that are holding back progress in the wireless sector. If their designs are promising, industry labs will pick them up and work out the implementation details. “As academic researchers, we realize that one or two PIs cannot work a concept from design all the way to market, particularly in our field,” Shen said. “We're taking good ideas from academic research papers and making them more useful.” Shen has leveraged his industry experience to find interesting problems that come from industry practice. “When I was in industry, I confronted research problems that a deeper, more profound theoretical impact was lacking. I had to set those problems aside to expedite products to market,” Shen said. “At UVA, I can adopt a general design philosophy to solve bigger problems to advance multiple applications simultaneously.”