Bayesian Tensor Models for Multiway Structural Data: A Theoretical Study and Applications

2016 Fiscal Year Final Research Report

• Project/Area Number: 15K15955
• Research Category: Grant-in-Aid for Young Scientists (B)
• Allocation Type: Multi-year Fund
• Research Field: Statistical science
• Research Institution: Institute of Physical and Chemical Research
• Principal Investigator: Zhao Qibin 国立研究開発法人理化学研究所, 脳科学総合研究センター, 研究員 (30599618)
• Project Period (FY): 2015-04-01 – 2017-03-31
• Keywords: Tensor Factorization / Missing Data / Tensor Completion / Robust Factorization

Outline of Final Research Achievements

In this project, we study the probabilistic model of tensor factorizations and their applications to EEG signals and computer vision. 1) We proposed a Bayesian robust CP tensor factorization with missing data, which can infer automatically the underlying CP rank and capture outliers effectively from partially observed data without tuning parameters. 2) We proposed a tensor completion method with smoothness constraints over latent factors, which is particularly useful for visual data. 3) We proposed an efficient algorithm for non-negative Tucker decomposition by employing low-rank approximations of the gradient. 4) We proposed a nonlinear tensor partial least squares algorithm for multi-block tensor regression. 5) We applied our proposed methods to EEG artifact removal, image/video denoisiong, and MRI denoising.

Free Research Field

情報学