Seminar: Towards Scalable Intelligence: from Generative and Physical AI to Future Paradigms
Yu Zeng
Research Scientist
NVIDIA
Friday, March 7
9:30 - 10:30AM
1100 Torgersen Hall
Abstract
As text-based visual generation continues to break new ground in content creation, my research pushes beyond the established paradigm in two critical directions: incorporating multimodal information beyond text inputs, and transcending the virtual domain to achieve physical intelligence. In this talk, I will present my recent projects in this pursuit. First, I will introduce novel methods for visual generation that effectively integrate both semantic and visual inputs at varying levels of detail. Then, I will demonstrate how generative models can be applied to physical AI, showcasing their potential as a digital twin of the real world and as a policy model by generating robot actions. Finally, I will discuss how insights from this convergence of generative AI and physical applications can inform broader AI research, establishing a roadmap for scalable advancement across multiple fields.
Biography
Dr. Yu Zeng is a Research Scientist at NVIDIA. Her research advances artificial intelligence through two primary directions: generative AI and label-efficient learning for computer vision. Her PhD research focused on generative models for visual synthesis with multimodal and hierarchical inputs. Her most recent research focuses on advancing embodied AI using deep generative models. Her research has been integrated into several products at NVIDIA, including NVIDIA Cosmos and NVIDIA Edify. Before joining NVIDIA, she obtained her PhD from Johns Hopkins University. Before that, she worked as a researcher at Tencent. She also worked with the research teams at Adobe during PhD and Master’s studies.