Neural ‘Architecture’ search
Speaker : 이진호 교수(연세대학교)
Jinho LEE is an assistant professor in the Department of Computer Science at Yonsei University. He received the Ph.D. degree in electrical engineering and computer science from Seoul National University in 2016, and was at IBM Research, Austin until 2019. His research interests include finding architectural and machine-learning solutions to system-related issues.
When we say ‘architecture’, what do you picture in your mind?
Most researchers in our field these days would think of neural network architectures, but until a few years ago it meant computer architectures. In today’s talk, I will introduce a neural network-based approach to searching for neural network architecture and computer architecture at the same time, which is often called the ‘co-exploration’ problem.
Often based on a reinforcement learning approach, one of the major problems of the co-exploration is often its huge search cost, which can be a few hundred hours, if not thousands. To mitigate such drawbacks,
I suggest formulating the problem in a differentiable way, so that the entire problem can be solved in a single pass.