Tensor Learning Team Seminar (Talk by Mingyuan Bai of The University of Sydney).

This is an online seminar. Registration is required.
【Tensor Learning Team】
【Date】2022/March/23(Wed) 10:00am-11:00am (JST)

【Speaker】 Mingyuan Bai of The University of Sydney

Title: Deep Tensor Learning for Tensorial Time Series Analysis

Abstract: The recent decade has witnessed increasing attention to tensorial time series such as videos and multi-relational social networks, where each time instance is a tensor, i.e., a multi-way array. They possess rich and complicated spatio-temporal information, whereas simply flattening them into vectors as in traditional methods creates the loss of spatial structural information, causes the large number of parameters and incurs the curse-of-dimensionality issue. Tensor decomposition methods are naturally designed to analyse spatial information in tensors and resolve the curse-of-dimensionality issue. Therefore, we have proposed several deep tensor learning models based on tensor decomposition to directly process tensorial time series, in order to exploit their complicated spatio-temporal information without the curse-of-dimensionality issue. Furthermore, our proposed models also demonstrate enhanced interpretability, robustness to irregular missing time instances and desired theoretical properties such as uniqueness of the solution and consistency. In addition, two of our proposed models successfully marry the control theory to deep learning. Empirically, our findings also reveal that tensorial time series analysis prefers deep tensor learning models to traditional deep learning models such as recurrent neural networks. For applications, our research work introduced in this talk is also one of the first to bring deep learning to international relation studies.