News

Fu receives NSF grant to study neural network energy improvements
October 5, 2021
By
Stephen Greenwell
Xin Fu, Ph.D., an associate professor of Electrical and Computer Engineering, received $499,999 in funding for her grant proposal, “Enabling On-Device Bayesian Neural Network Training via An Integrated Architecture-System Approach.”
Xin Fu, Ph.D., an associate professor of Electrical and Computer Engineering, received $499,999 in funding for her grant proposal, “Enabling On-Device Bayesian Neural Network Training via An Integrated Architecture-System Approach.”

A Cullen College of Engineering professor has received just shy of $500,000 from the National Science Foundation to study how to make the decision-making networks in devices like self-driving cars and medical imaging devices more efficient.

Xin Fu, Ph.D., an associate professor of Electrical and Computer Engineering, received $499,999 in funding for her grant proposal, “Enabling On-Device Bayesian Neural Network Training via An Integrated Architecture-System Approach.”

Fu described her research as looking at ways to improve the energy needs of products that rely on machine learning, which is one of her research interests.

“This project targets at energy-efficiently training Bayesian Neural Networks (BNNs) locally on mobile devices,” she said. “BNNs are generally used in real-world AI applications that request reliable and robust decision-making, like autonomous driving and medical image diagnosis. In this project, we will explore hardware-software co-designed methods to significantly reduce the computation workloads and memory accesses required by BNN training. This will enable the BNN training on the resource-constrained mobile devices.”

Fu stressed that finding more energy-efficient ways to train BNNs was vital for improving their reliability and safety.

“Deep Neural Networks have recently achieved amazing success in numerous AI applications. However, the neural network models can become unreliable due to the uncertainty in data, giving a false judgement and incurring a disaster,” she said. “To address this issue, BNNs have been increasing applied in a wide range of real-world AI applications, which demand reliable and robust decisions. Training a BNN model is very time and energy consuming, so this motivates us to explore the energy-efficient training for BNN models. Enabling the BNN training on mobile devices, and furthermore, at the edges, can greatly help the AI technologies become more adaptive and intelligent to fast-changing environments, and benefit every day living and working.”

Fu identified one doctoral student – Qiyu Wan – as already working on the research, but she is interested in hiring more. For more information, visit the prospective students page of her Efficient Computer Systems (ECOMS) Lab. Research on this grant started Oct. 1, and is estimated to last three years.

Faculty

Share this Story: