Dingwen Tao, associate professor of intelligent systems engineering at the Luddy School of Informatics, Computing and Engineering, has received a 2022 Meta Research Award for $50,000.
Meta received 62 proposals from 47 universities around the world for its AI System Hardware/Software Codesign request for proposals. As lead investigator, Tao partnered with Tony Geng of the University of Rochester on the proposal, “Accelerating communication in DLRM via frequency-aware lossy compression.” It was one of just nine honored.
“It is a great honor to receive this Meta Research Award on AI System Hardware/Software Codesign,” Tao said. “It takes our research and software from the scientific community to the industrial world. I am excited to partner with Meta Research to explore how our compression technology can benefit production computing systems for world-class machine learning and artificial intelligence workloads.”
Meta AI teams use codesign to develop high-performance AI solutions for existing and future AI hardware. It has reached out to institutions to boost innovation and to support academics seeking to further explore codesign opportunities across a number of new dimensions.
“Dingwen is known for pursuing innovative research that elevates the understanding of what is possible and translates that to real-world use,” said Martin Swany, chair of ISE. “Being recognized by a tech giant such as Meta showcases the Luddy School’s difference-making faculty. It will facilitate future collaborations between Indiana University, the Luddy School and Meta. We are lucky to have him.”
Tao said Meta seeks to better determine which advertisements people will view by using Artificial Intelligence such as Deep Learning Recommendation Model to predict click-through rate. He said Meta is the leading company to do personalization recommendation, and that when you look at Facebook, it tries to target you with advertising it thinks you’ll like.
“If I buy something, maybe my friends or co-workers will buy it. Facebook wants to promote items to the same group of people with similar features.”
Tao said Facebook wants to know the chances that people will click on an advertisement and whether or not it should send advertisements to a certain group of people. The click-through rate is the key.
The DLRM needs constant updating. Doing that requires “a lot of time, resources and effort,” Tao said. It involves the use of hundreds of high-performance computing servers. Transmitting data to all those servers is time consuming.
“We can help accelerate the process and reduce the training time by using compression technique,” Tao said.
He added it takes about 10 hours of training, including an hour or two of data transmission, to update the DLRM. For larger terabyte datasets, each training could take a few weeks. There also can be long pre-process before each training session. The proposal could cut training time in half.
“We have built a very good compression software that can compress the data a lot and still maintain good information.,” Tao said. “Meta will fund us to help reduce the amount of data transferred between different computing servers.”