Neural networks inspired by differential equations have proliferated over the past few years. Neural ordinary differential equations (NODE) and neural control differential equations (NCDE) are two representative examples. In this seminar, I will introduce a study on applying attention to NCDEs, namely Attentive Neural Controlled Differential Equation (ANCDE), which uses dual NCDE modules. One is to generate attention values, and the other is to evolve hidden vectors, referring to those attention values, for a downstream machine learning task. This work conducted experiments using three real-world time-series datasets and ten baselines. This method consistently shows the best accuracy in all cases. I will also introduce interesting visualization outcomes to show how the presented attention mechanism works.
Sheo Yon Jhin is a first-year PhD student in the Department of Artificial Intelligence at Yonsei University, Seoul, South Korea. Her main research interest is time-series classification and forecasting based on differential equation-based models. She has published three papers in AAAI 2021, KDD 2021, and ICDM 2021.