2019 Fiscal Year Annual Research Report
High-Order Deep Learning Models: Theoretical Study and Applications
| Project/Area Number |
17K00326
|
|---|
| 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, 革新知能統合研究センター, ユニットリーダー (30599618)
|
|---|
| Co-Investigator(Kenkyū-buntansha) |
曹 建庭 埼玉工業大学, 工学部, 教授 (20306989)
|
|---|
| Project Period (FY) |
2017-04-01 – 2020-03-31
|
| Keywords | Tensor Ring / Deep Multimodal Learning |
|---|
| Outline of Annual Research Achievements |
We proposed a robust tensor decomposition approach for tensor completion by defining a new tubal nuclear norm on tensors. The tubal nuclear norm has orientation invariant property which is the key contribution to improve the robustness and performance for tensor completion task.
We developed a deep multimodal learning model based on multilinear tensor fusion of latent layers. The proposed tensor fusion framework is able to capture high-order interactions for inter-modal and intra-modal features, which has more expressive power. In addition, we apply tensor network to represent weight parameters, which thus reduce the computation complexity dramatically. The proposed method can improve the performance while not increasing model and computation complexity.
We studied the theory analysis of matrix completion under linear transformations. This work provides rigorous theory support for many "non-local" based low-rank completion methods. In addition, the proposed framework is able to improve matrix completion performance by low-rankness under the multiple transformations. Experiments results show its advantages in image in-painting task.
We have developed several tensor completion algorithms based on tensor ring model. By defining tensor ring based nuclear norm, we can solve low-rank tensor approximation by nuclear norm optimizations. For large-scale data, we have developed an efficient randomized tensor ring decomposition algorithm, which is fast and scalable to very large tensors.
|
