Two PhD students from the Ming Hsieh Department of Electrical and Computer Engineering and the Department of Computer Science, Chaoyang He and Saurav Prakash, advised by Prof. Salman Avesttimehr, have recently won prestigious Qualcomm Innovation Fellowships. The students’ proposal, titled “Federated Deep Learning: On-device Learning of CV and NLP with Transformers and CNNs,” was one of only 16 proposals accepted. Of those accepted proposals, Prakash and He were one of only two teams to turn in work in the new field of federated learning. This research area is vital to build more trust and security in the AI and machine learning systems that play an increasingly important role in society.
“Machine learning requires large data sets to produce an effective model, Prakash said. “In the past, when accessing such large amounts of data, personal privacy has been inadvertently compromised, and this has led to a deep distrust of AI and machine learning in society.”
Imagine a federation of hospitals working together to understand a new disease. Each individual hospital has detailed information about many patients. If they could combine their data, they could train a machine learning system to better understand the disease. Unfortunately, this would be a serious invasion of patient privacy and also a security risk.
“Federalized learning could allow these hospitals to securely build a dataset of their combined information without actually sharing the information with each other,” He said. In a world where more effective and trusted AI systems are becoming increasingly important to society, this progress cannot be underestimated.
Prakash and He, both members of Professor Salman Avesttimehr’s research group, have been working on this challenge with their advisor and fellow students for some time. The team has even established an open source platform, fedml.ai, to enable research and development on federated learning across a variety of application domains. The FedML ecosystem, which has already attracted more than 600 active users worldwide and is in the top 3 GitHub repositories for federated learning, provides a comprehensive scientific platform to help researchers around the world build truly secure and reliable AI and machine learning.
With this fellowship, the duo plans to improve the efficiency of federated learning programs as a whole so that the technology can be implemented on a larger scale. “This fellowship provides us with an excellent opportunity to build stronger, long-term partnerships with Qualcomm in the field of federated learning,” said Avesttimehr.