2015 Fiscal Year Research-status Report
Bayesian Tensor Models for Multiway Structural Data: A Theoretical Study and Applications
| Project/Area Number |
15K15955
|
|---|
| 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 | Machine Learning / Tensor Factorization / Image Completion / Image Denoising |
|---|
| Outline of Annual Research Achievements |
I have proposed and developed a Bayesian model for multiway tensor decomposition. Our method can use incomplete data as input, and provide the completion of whole data. In addition, the rank of tensor model can be automatically inferred. Our method has been used to image completion and achieve the best performance in the world.
I have proposed and developed Bayesian robust tensor factorization method for data which contains outliers or non-Gaussian noise. This method can detect sparse noise effectively. We have applied this method to image denoising, and background and foreground separation of videos, which achieve the impressive performance.
I have also proposed Bayesian model for Tucker-style tensor factorization, which can also deal with missing data case. The method can be used for EEG completion, and feature extraction of multiway data.
I have applied above methods for variety of applications, such as image denoising, video analysis, EEG classification and MRI completions. The experimental results demonstrate the effectiveness of our methods.
|
| Current Status of Research Progress |
Current Status of Research Progress
2: Research has progressed on the whole more than it was originally planned.
Reason
The project was performed smoothly. We have not only finished the initially planned work, but also performed several new applications using our methods. These are motivated during my research procedure.
|
| Strategy for Future Research Activity |
In the next fiscal year. My research plan includes that
1. The kernel technique will be employed in tensor decomposition methods, which may provide many advantages.
2. Tensor network such as tensor train will be investigated from Bayesian point of view. We aim to develop a Bayesian tensor train method.
3. The big data or big tensor problem will be also considered. We plan to use tensor network to handle the big data problem to some extend.
|
| Causes of Carryover |
The several related international conference will be held in this year. The first year, I focus on the theoretical research. The real world applications have less studied, which makes EEG recording delayed.
|
| Expenditure Plan for Carryover Budget |
I will employ a Ph.D student in this year to perform real world applications of tensor methods. We will also consider to attend several international conference and give the oral presentations.
|
