Fifth-year PhD candidate Bojian Zheng (Photo: Supplied)
Two computer science PhD students are among an illustrious group of researchers who have been awarded a 2023 Qualcomm Innovation Fellowship.
Yaoyao Ding and Bojian Zheng comprise one of 18 winning teams from among 43 finalists and the only students from a Canadian university to be selected this year.
Their work aims to deploy diverse artificial intelligence models with dynamic structures (e.g., large language models) into modern accelerators (e.g., NVIDIA GPUs) efficiently.
“The goal of this project is to improve the runtime performance of key machine learning applications on modern computer platforms such as GPUs,” says Zheng, a fifth-year PhD candidate and whose research focuses on efficient computer system design.
“The fellowship is a recognition of both Yaoyao’s and my research contributions in the area. It motivates us to pursue solving real-world problems that are important for academia and industry.”
“I am quite excited upon receiving the fellowship. In addition to financial support, the fellowship clearly tells us that we are working on a significant problem and proposing an interesting and compelling solution,” he adds.
Their winning proposal, “Dynamic Deep Neural Network Compilation” was selected from a highly competitive pool of over 100 submissions.
In this project, Ding and Zheng propose Hetron, a holistic system for compiling dynamic Deep Neural Networks (DNNs). It seeks to address the current challenge that existing compilers are limited to static input models, hindering support for dynamic DNNs.
“Hetron decouples condition and loop constructs from the computation graph and defines schedule templates with symbolic input shapes for on-demand specialization. This allows Hetron to efficiently handle dynamic execution flows and input sizes, outperforming existing deep learning frameworks,” the team states in their proposal abstract.
Second-year PhD student Yaoyao Ding (Photo: Supplied)
“I am thrilled that my research work has been recognized by researchers in the industry. The fellowship helps me collaborate with Qualcomm researchers and apply our research on mobile devices,” says Ding, a second-year PhD student focusing on efficient machine learning systems.
Ding and Zheng are supervised by Associate Professor Gennady Pekhimenko.
“This breakthrough research allows us to radically improve the speed and reduce the cost of modern machine learning models, including the ones used for popular applications such as ChatGPT. It is only the second time ever that someone from the University of Toronto has won this prestigious award,” says Pekhimenko.
The Qualcomm Innovative Fellowship program run by the wireless and mobile technologies company focuses on “recognizing, rewarding, and mentoring innovative PhD students across a broad range of technical research areas,” and “supports them in their quest towards achieving their research goals,” the website reads.
The QIF program is open to students from a preselected variety of top U.S.-based and international schools. Winning students earn a one-year fellowship and are mentored by Qualcomm’s engineers to facilitate the success of the proposed research.