Su Receives NSF CAREER Award to Strengthen Federated Learning
Lili Su, assistant professor, electrical and computer engineering (ECE), has been awarded a $611,000 National Science Foundation (NSF) CAREER grant to consolidate the theoretical foundations and enrich the algorithmic toolbox of distributed machine learning.
Federated learning is a communication-efficient distributed machine learning approach that enables training global models without sharing raw local data. Used in commercial applications such as autonomous vehicles, internet of things, industrial automation, and more, federated learning has seen a recent upswing due to increased demand for faster data processing and better privacy.
Widely implemented distributed machine learning algorithms have faced recent criticism that they don’t work as well as they should. Su’s research seeks to quantify the effectiveness of existing algorithms and design new and more efficient ones, all with a focus on enhancing resilience against three main challenges: data heterogeneity, inherent system faults, and external attacks.
“Machine learning relies heavily on the value of its data, and those informational structures are not well taken care of in existing theoretical predictions,” says Su. “One component of my proposal is to bridge the gap between existing theory and practice by examining the underlying statistical structure of the datasets.”
Another component of Su’s research is to address an innate problem of distributed systems: Because they work in an open environment, they are prone to system imperfections.
“In the example of connected and autonomous vehicles, you have different agencies with different computational powers, as well as different vehicle owners driving at different times to different places—it’s a very complex environment,” says Su. “The algorithm used by the system needs to resilient to such differences, as well as a wider range of uncertainties.”
A third component is increasing resilience against adversarial scenarios in which calibrated attacks seek to compromise or manipulate certain agents in the distributed system.
“Overall, this research proposal looks at a wide spectrum of problems, from theoretical analysis on one end to casting algorithms on real-world data and cyber-physical systems on the other,” says Su. “I’m interested in the larger potential of this work across applications.”
In addition to the impact of the research on distributed systems across industries, Su also intends to use her CAREER grant to expand education in machine learning toward broader and younger audiences.
“With the educational component, we’re targeting female/woman-identifying students and under-represented minority groups in K–12,” says Su. “Students in higher grades will be involved in projects to see how algorithms can be implemented in real-world contexts like connected and autonomous vehicles. For younger students, we will create mini-courses that illustrate the power of distributed system solutions, as well as the importance and usefulness of mathematics and rigorous reasoning.”
This last piece speaks to Su personally; as a young student, she recalls not seeing the value of mathematics outside of just puzzles and games.
“As I grew up and began doing more research, I truly saw the power of mathematics and its ability to help us understand things and construct solutions, especially with complicated issues,” says Su. “I want to help young students learn the mathematical tools they need to help them navigate challenges and do rigorous reasoning on big problems, now and in the future.”
NSF Source: NSF
Related Article: Lili Su receives NSF CAREER Award for developing ‘resilient, scalable distributed algorithms’