This talk explores the principles and challenges of U-Net, a foundational architecture in image processing. The first segment focuses on its mathematical foundations, explaining how the encoder-decoder framework enables efficient learning of sparse structural features by analyzing convolutional operations and transposed convolutions. The second segment addresses hallucination artifacts—plausible but inconsistent structures generated during inference. We present a distribution-free uncertainty quantification framework designed to quantitatively evaluate hallucination severity, emphasizing its potential for network architecture choices. Experimental comparisons between U-Net and SUNet, trained with L1 and L2 losses, reveal how variations in architecture and loss functions influence the nature of artifacts.
Jianfei Li received the B.S. degree in mathematics from Ocean University of China, Qingdao, China, in 2017, the M.S. degree in mathematics from Sun Yat-sen University, Guangzhou, China, in 2020, and the Ph.D. degree in mathematics from City University of Hong Kong, Hong Kong, China. He is currently a Post-Doctoral fellow with the LMU Munich, Munich, Germany collaborating with Prof. Gitta Kutyniok. His research interests include approximation theory, signal processing, and deep neural networks.Jianfei Li received the B.S. degree in mathematics from Ocean University of China, Qingdao, China, in 2017, the M.S. degree in mathematics from Sun Yat-sen University, Guangzhou, China, in 2020, and the Ph.D. degree in mathematics from City University of Hong Kong, Hong Kong, China. He is currently a Post-Doctoral fellow with the LMU Munich, Munich, Germany collaborating with Prof. Gitta Kutyniok. His research interests include approximation theory, signal processing, and deep neural networks.
