A Novel One-shot Neural Network Architecture Search via Regression
Neural architecture search (NAS) has become a central theme for the recent research trend on autoAI, i.e., how to find an effective deep neural network (DNN) by automatically searching a large neural architecture design space. One major drawback is, however, the large computation time used to search a high-quality neural architecture. In this talk, we will discuss one of our recent works in NAS where we show a simple intuition that was surprisingly effective in guiding us to design a fast NAS, which has achieved new state-of-the-arts results in both image and natural language related tasks but with orders of magnitude performance speedup. We call this technique as Generic Neural Architecture Search via Regression (GenNAS). If time permits, I will also discuss how we ended up doing NAS research in this space, and what new theoretical research topics may be needed in order to better understand GenNAS for its generalization capabilities.
Dr. Jinjun Xiong is an Empire Innovation Professor with the Department of Computer Science & Engineering, University at Buffalo (UB). He received his Ph.D. degree in 2006 from University of California, Los Angeles (UCLA) with an Outstanding Ph.D. Award, his M.S. degree from University of Wisconsin, Madison in 2002, and his M.S. and B.S. degrees from Tsinghua University in 2000 and 1998, respectively. Before joining UB in 2021, Dr. Xiong was Program Director and Senior Research Scientist at IBM T.J. Watson Research Center, Yorktown Heights, NY. He co-founded and co-directed the IBM-Illinois Center for Cognitive Computing Systems Research with Prof. Wen-mei Hwu. Under their leadership, the C3SR center has expanded from the early days' eight faculty members in 2016 to close to 40 faculty members in 2021. The success of the C3SR center also led to the creation of the new IBM-Illinois Discovery Accelerator Institute, a joint $200-million research investment between IBM and UIUC. Dr. Xiong also co-founded the IBM Smarter Energy Research Institute and led a number of enterprise-scale collaborations with world-wide electric utility companies to address sustainability issues with renewable integration.
He has published more than 150 peer-reviewed papers in top AI conferences and systems conferences. His publication has won seven Best Paper Awards and eight Nominations for Best Paper Awards. Dr. Xiong also won top awards from various international competitions, including the recent Championship Award for the IEEE GraphChallenge on accelerating sparse neural networks, and the Championship Awards for the DAC'19 Systems Design Contest on designing an object detection neural network for both edge FPGA track and the edge GPU track. Many of his research results have been adopted in commercial enterprise-scale products and tools.
27/06/2022 14:00
Knowledge-Based Systems
Dr. Jinjun Xiong is an Empire Innovation Professor with the Department of Computer Science & Engineering, University at Buffalo (UB). He received his Ph.D. degree in 2006 from University of California, Los Angeles (UCLA) with an Outstanding Ph.D. Award, his M.S. degree from University of Wisconsin, Madison in 2002, and his M.S. and B.S. degrees from Tsinghua University in 2000 and 1998, respectively. Before joining UB in 2021, Dr. Xiong was Program Director and Senior Research Scientist at IBM T.J. Watson Research Center, Yorktown Heights, NY. He co-founded and co-directed the IBM-Illinois Center for Cognitive Computing Systems Research with Prof. Wen-mei Hwu. Under their leadership, the C3SR center has expanded from the early days' eight faculty members in 2016 to close to 40 faculty members in 2021. The success of the C3SR center also led to the creation of the new IBM-Illinois Discovery Accelerator Institute, a joint $200-million research investment between IBM and UIUC. Dr. Xiong also co-founded the IBM Smarter Energy Research Institute and led a number of enterprise-scale collaborations with world-wide electric utility companies to address sustainability issues with renewable integration.
He has published more than 150 peer-reviewed papers in top AI conferences and systems conferences. His publication has won seven Best Paper Awards and eight Nominations for Best Paper Awards. Dr. Xiong also won top awards from various international competitions, including the recent Championship Award for the IEEE GraphChallenge on accelerating sparse neural networks, and the Championship Awards for the DAC'19 Systems Design Contest on designing an object detection neural network for both edge FPGA track and the edge GPU track. Many of his research results have been adopted in commercial enterprise-scale products and tools.
