High risk research results in reward for former CSL student
Former ITI student Mingu Kang will receive the Coordinated Science Laboratory's (CSL) Ph.D. Thesis award this month for his dissertation “Deep In-memory Computing” for realizing energy and latency efficient machine learning systems in silicon. The award is given to one student every calendar year whose thesis best aligns with CSL’s goals of interdisciplinary collaboration and making an impact on science.
“There is a huge variety of work that goes on here so the competition is very high,” said Bruce Hajek, award committee chairman and professor in the Department of Electrical and Computer Engineering. “Winning this award is a real achievement.”
Kang will be the third student to receive the award, and the recognition comes after a long and arduous process. He was encouraged to think in terms of “high risk, high gain” research by his adviser Naresh Shanbhag, which, while ultimately successful, wasn’t easy. Many of Kang’s papers didn’t get accepted for publication until after he graduated.
“Professor Shanbhag doesn’t like to try and chase low-hanging fruit,” said Kang, now an ECE graduate. “My Ph.D. was high risk, but at the time I didn’t get high gain yet. It was very painful to open new branch of research area and demonstrate the novel idea in real silicon hardware. This CSL award symbolizes a turning point in my work finally receiving recognition.”
In his research, Kang conceived of the deep in-memory architecture, or DIMA, for reducing the energy consumption of machine learning systems used in battery-powered electronics.
“DIMA is a radical shift away from von Neumann architecture since it eliminates the standard memory-processor interface,” said Shanbhag. “Mingu was extremely bold in not only formulating DIMA but also working hard to make it real by designing multiple prototype chips.”
The IC prototypes that Kang and his colleagues designed in 65nm CMOS technology has demonstrated up to 100 times lower energy-delay product than today’s digital systems.
While it may have taken awhile, Kang’s efforts were noticed in and out of CSL. His research leading up to his dissertation was innovative, looking at the effectiveness of low precision analog computing of machine learning algorithms, a topic not many others had thought about. In addition to numerous publications, Mingu and Shanbhag were awarded a US patent on DIMA titled “Compute Memory” in 2017. Since his initial studies, the amount of literature on the topic has increased drastically, even leading to an entire session devoted to the idea at the most recent Institute of Electrical and Electronic Engineers (IEEE) International Solid-State Circuit Conference, as well as an IEEE Spectrum article.
This impact on his field of study was an important aspect of the application for the award.
“Mingu’s work has influenced major semiconductor research programs such as JUMP and ERI where analog in-memory architectures have been called out,” said Shanbhag. “Researchers in major semiconductor companies and universities are following up on this work. The impact of Mingu’s CSL dissertation will be felt for many years to come.”
Also important were his collaborations with researchers in computer science, circuits, architecture and theory - a cooperation Kang says was largely impacted by being in CSL.
“My research is unique in that it is very well integrated across different fields of study including circuit, architecture, compiler, systems, and theory,” Kang said. “It’s not really common and was definitely influenced by the CSL environment and research.”
After graduating with a degree in electrical and computer engineering in 2017, Kang joined the IBM TJ Watson Research Center in New York as a research staff member. His current work is an extension of his Ph.D. research; he applies his learning to real-life products by optimizing the data processing and movement in large-scale machine learning processors. Kang will be returning to CSL on September 17 to receive the award and give a presentation about his winning-research.