Yingjie Li

Research Abstract:

The era we are in is an energetic time for eager calls for powerful computing approaches in both algorithm and new infrastructure as we have seen the limitations in Moore's law, which impedes the development of digital systems. More and more people turn their eyes to another information carrier -- light signal, which features with its high computation speed and rich information encoding dimensions, which shows the potential to deal with the explosive data-related tasks with rich feature embedding and parallelism computing. On the other side, we also find the importance of developing EDA tools under the current international competitions. As a result, those motivate my research in two directions: Optical AI and hardware synthesis. Optical AI -- Optical neural network is very promising in improving computing efficiency and saving energy for computation-intensive tasks. However, the development of optical computing is limited due to the high exploration technical barrier. To make it easier for researchers to explore optical computing, we first proposed the first end-to-end agile design framework LightRidge for diffractive optical neural networks, which implements a physics-aware co-design algorithm designed specifically for DONN, enabling fast, accurate and efficient optical neural network emulation and deployment. Based on the framework, we (1) developed the first adversary analysis with DONN system in optical domain with the prototype demonstration; (2) proposed the physics-aware multi-task learning for the DONN system ; (3) explored the interpixel interaction problem in the physical implementation of DONNs and provided the algorithm to migrate the problem. Hardware synthesis -- Synthesis is an important step in EDA design flow and makes a profound influence on the final performance. My research has mainly focused on developing novel fundamental optimization algorithms for end-to-end design closure. Furthermore, we took advantage of machine learning in dealing with logic synthesis, including (1) FlowTune, which using multi-armed bandits to optimize the logic optimization sequence for better optimization performance; (2) SLAP, which is believed to be the first work using ML model to find the optimal cut for technology mapping; (3) Gamora, which employs graph learning for fast and accurate structure extraction and word-level reasoning ; (4) RESPECT, which employs reinforcement learning to produce sequential scheduling; (5) DAG-aware orchestration that offers new perform frontier over classic logic synthesis approaches; (6) BoolGebra, which optimizes logic synthesis in a differentiable way with machine learning for customized optimization targets. Future research -- I plan to contribute to more broad communities with cross-disciplinary research efforts in both optical computing and EDA, with focuses on physics-based AI systems, optical design automation, and ML for EDA and generic optimization.

Bio:

Yingjie Li is pursuing Ph.D. at the University of Maryland, College Park, under the supervision of Prof. Cunxi Yu, starting from 2020 Spring. She received B.S. degree in 2018 from Huazhong University of Science and Technology in China, and her Master degree from Cornell University in 2019. Her recent research interests focus on hardware-software codesign and compilation systems for optical neural networks and physics-aware adversarial machine learning, and machine learning for EDA. She joined NVIDIA in the summer of 2023 for the internship. Yingjie was selected as Design Automation Conference (DAC) Young Fellow in 2020, 2021, and 2022, and received Best Paper Presentation Award in 2020, Best Paper Award at DAC'23, and Rising star in 2023.