Abstract
Achieving reliable and robust performance of visual perception systems with Artificial Intelligence (AI) in real-world industrial applications can be challenging due to highly dynamic conditions and requirements. Traditional AI-based solutions require offline training on high-quality data, making it difficult to accommodate new technologies, data, hardware, tasks, and customer needs without retraining. To address this issue, Dynamic Deep Learning is proposed, which utilizes condition-invariant properties, dynamic architectures, and dynamic training strategies to develop ML solutions that can adapt to new operational environments and requirements in a fast, reliable, and cost-effective manner. This presentation focuses on three sub-problems: dynamic environments, dynamic modalities, and dynamic tasks. Dynamic environments and modalities cover the input data and sensing modality challenges, while dynamic tasks cover the output task challenges. The talk highlights my research efforts aimed at solving these issues, emphasizing the potential of Dynamic Deep Learning to provide reliable, consistent, and robust performance for real-world deployment.
Presenter
Dr. Yuhao Chen is currently a postdoctoral fellow at the Vision and Image Processing Lab (VIP) in Systems Design Engineering at the University of Waterloo. He obtained his B.A.Sc. and Ph.D. degrees in Electrical and Computer Engineering from Purdue University in 2015 and 2019, respectively, while being a member of the Video and Image Processing (VIPER) laboratory. With a strong passion for exploring cutting-edge technologies, Yuhao's research has been focused on developing computer vision and artificial intelligence solutions for industrial applications, specifically in the fields of precision agriculture, precision health, and precision manufacturing.