Dr. Sun Yuan Kung

Biography

Sun-Yuan Kung's research focuses on developing high-performing learning networks and deep learning processors that have effectively replaced traditional signal processors in speech and image processing applications. His work on back-propagation, a critical training paradigm for deep learning models, emphasizes the importance of finding optimal network structures. Through structural learning, he highlights node and layer importance and proposes effective mechanisms such as node-ranking and layer ranking to facilitate network pruning and deep compression. His innovative approach includes a joint parameter/structural gradient-type method that polishes networks towards optimal configurations by exploring comprehensive solution spaces for design parameters including size, power, speed, and accuracy. Additionally, Kung has developed internal learning paradigms that utilize internal teacher labels and internal optimization metrics, leading to the creation of Explainable Neural Networks (XNN) that enhance model robustness by removing redundant nodes based on discriminant information. His research work is pivotal in advancing machine learning, particularly in making neural network operations comprehensible and effective in dynamic environments such as active learning scenarios.

Research Interests