| Presentation | 2020-03-17 Deep neural network representation and learning of low-rank and sparse approximation Ryohei Miyoshi, Tomoya Sakai, Takashi Ohnishi, Hideaki Haneishi, |
|---|---|
| PDF Download Page |  PDF download Page Link |
| Abstract(in Japanese) | (See Japanese page) |
| Abstract(in English) | Low-rank and sparse (L+S) approximation, a.k.a. stable and robust principal component analysis, is known to be suitable for analyzing sequential data simultaneously containing linearly dependent and sparse features. There remain, however, some challenges in practice. Hyperparameters ${bf Lambda}_{mbox{scriptsize S}}$ to control the balance between the low-rankness and sparseness must be tuned on real data. Introducing additional prior knowledge to the L+S structures makes the optimization algorithm computationally intensive. This work takes a novel approach to improve the L+S approximation by representing its algorithm as a deep neural network (DNN) with trainable hyperparameters ${bf Lambda}_{mbox{scriptsize S}}$. In order for the DNN to learn via the backpropagation, the singular-value thresholding and soft thresholding in the L+S algorithm, as well as the nuclear and $ell_1$ norms imposing the L+S nature on the DNN outputs, can be implemented as auto-differentiable modules using a deep learning framework. Introducing the total variation of the sparse components into the loss function for the DNN training, the hyperparameters successfully acquire spatial distribution of the sparse components, which is experimentlly shown to improve artery detction in the application to celiac angiography under free-breathing condition. |
| Keyword(in Japanese) | (See Japanese page) |
| Keyword(in English) | Robust PCA / Hyperparameter optimization / Nuclear loss function / ADMM |
| Paper # | PRMU2019-91 |
| Date of Issue | 2020-03-09 (PRMU) |
| Conference Information | |
| Committee | PRMU / IPSJ-CVIM |
|---|---|
| Conference Date | 2020/3/16(2days) |
| Place (in Japanese) | (See Japanese page) |
| Place (in English) | |
| Topics (in Japanese) | (See Japanese page) |
| Topics (in English) | |
| Chair | Yoichi Sato(Univ. of Tokyo) |
| Vice Chair | Toru Tamaki(Hiroshima Univ.) / Akisato Kimura(NTT) |
| Secretary | Toru Tamaki(NTT) / Akisato Kimura(OMRON SINICX) |
| Assistant | Yusuke Uchida(DeNA) / Takayoshi Yamashita(Chubu Univ.) |
| Paper Information | |
| Registration To | Technical Committee on Pattern Recognition and Media Understanding / Special Interest Group on Computer Vision and Image Media |
|---|---|
| Language | JPN |
| Title (in Japanese) | (See Japanese page) |
| Sub Title (in Japanese) | (See Japanese page) |
| Title (in English) | Deep neural network representation and learning of low-rank and sparse approximation |
| Sub Title (in English) | With application to celiac angiography under free breathing |
| Keyword(1) | Robust PCA |
| Keyword(2) | Hyperparameter optimization |
| Keyword(3) | Nuclear loss function |
| Keyword(4) | ADMM |
| 1st Author's Name | Ryohei Miyoshi |
| 1st Author's Affiliation | Nagasaki University(Nagasaki Univ.) |
| 2nd Author's Name | Tomoya Sakai |
| 2nd Author's Affiliation | Nagasaki University(Nagasaki Univ.) |
| 3rd Author's Name | Takashi Ohnishi |
| 3rd Author's Affiliation | Chiba University(Chiba Univ.) |
| 4th Author's Name | Hideaki Haneishi |
| 4th Author's Affiliation | Chiba University(Chiba Univ.) |
| Date | 2020-03-17 |
| Paper # | PRMU2019-91 |
| Volume (vol) | vol.119 |
| Number (no) | PRMU-481 |
| Page | pp.pp.133-138(PRMU), |
| #Pages | 6 |
| Date of Issue | 2020-03-09 (PRMU) |