Title：Genealized constraints for knowledge-driven-and-data-driven approaches
Speaker: Baogang Hu
Abstract: When the Deep Learning(DL), as a data-driven approach, is successful in many application areas, we believe that knowledge-and-data-driven modeling(KDDM) approaches will be the next for advancing the existing tools. However, the current Artificial Neural Networks(ANNs), including DL, are not ready to incorporate any types of prior knowledge. They are considered as “black box” tools. For overcoming the difficulties, we adopt the notion of “Generalized Constraints(GCs)” in modeling. Different with the conventional constraints, like well defined equality and inequity, GCs appear more often inapplications, which may show unstructured or even partially known prior information. We show that GCs will enlarge our modeling tasks, such as coupling forms in KDDM, guaranteed performance over the ANNs after embedding GCs, identifiability of parameters, etc. Examples are given on regression and dynamic process problems. The main objective of this talk is to highlight GCs from a KDDM perspective in relation to novel mathematic challenges, rather than to specific applications.
Baogang Hu, Professor, Senior Member, IEEE
National Laboratory of Pattern Recognition
Institute of Automation
Chinese Academy of Sciences, Beijing China
Baogang Hu received his Ph.D. degree in 1993 from Department of Mechanical Engineering, McMaster University, Canada. Currently, he is a professor of National Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences (CAS), Beijing, China. From 2000 to 2005, he served as the Chinese Director of “Chinese-French Laboratory of Information, Automation and Applied Mathematics”(LIAMA) . His current researches include machine learning and plant growth modeling.
|Date||August 29, 2019 (Thu) 14:00 - 15:00|