| Presentation | 2018-03-26 Suppression Method of Mode Collapse in Generative Adversarial Nets Shinya Hidai, Hidehiro Nakano, Arata Miyauchi, |
|---|---|
| PDF Download Page |  PDF download Page Link |
| Abstract(in Japanese) | (See Japanese page) |
| Abstract(in English) | Generative Adversarial Nets (GAN) is constituted by two neural networks, Generator and Discrminator. Generator creates data, and Discriminator estimates data that came from training data. The purpose of GAN is to let Generator estimate the probability distribution of training data. There is a problem in the GAN that the Generator can only estimate the local maximum of a part of the probability distribution of the training data, which is known as Mode Collapse. In this paper, in order to suppress the occurrence of Mode Collapse, we propose a method to change loss function of Generator to an expression with a large gradient even when learning of Generator proceeds. We verify this proposed method by estimating the Gaussian mixture model in two dimensional space. |
| Keyword(in Japanese) | (See Japanese page) |
| Keyword(in English) | Generative Adversarial Nets / Mode Collapse / probability distribution / loss function |
| Paper # | CCS2017-33 |
| Date of Issue | 2018-03-19 (CCS) |
| Conference Information | |
| Committee | CCS |
|---|---|
| Conference Date | 2018/3/26(2days) |
| Place (in Japanese) | (See Japanese page) |
| Place (in English) | Tokyo Univ. of Sci. (Morito Memorial Hall) |
| Topics (in Japanese) | (See Japanese page) |
| Topics (in English) | Natural Computing, etc. |
| Chair | Naoki Wakamiya(Osaka Univ.) |
| Vice Chair | Mikio Hasegawa(Tokyo Univ. of Science) / Makoto Naruse(NICT) |
| Secretary | Mikio Hasegawa(Osaka Univ.) / Makoto Naruse(Tokyo City Univ.) |
| Assistant | Chisa Takano(Hirishima City Univ.) / Takashi Shimada(Univ. of Tokyo) / Tomoya Suzuki(Ibaraki Univ.) / Ryo Takahashi(AUT) |
| Paper Information | |
| Registration To | Technical Committee on Complex Communication Sciences |
|---|---|
| Language | JPN |
| Title (in Japanese) | (See Japanese page) |
| Sub Title (in Japanese) | (See Japanese page) |
| Title (in English) | Suppression Method of Mode Collapse in Generative Adversarial Nets |
| Sub Title (in English) | |
| Keyword(1) | Generative Adversarial Nets |
| Keyword(2) | Mode Collapse |
| Keyword(3) | probability distribution |
| Keyword(4) | loss function |
| 1st Author's Name | Shinya Hidai |
| 1st Author's Affiliation | Tokyo City University(Tokyo City Univ.) |
| 2nd Author's Name | Hidehiro Nakano |
| 2nd Author's Affiliation | Tokyo City University(Tokyo City Univ.) |
| 3rd Author's Name | Arata Miyauchi |
| 3rd Author's Affiliation | Tokyo City University(Tokyo City Univ.) |
| Date | 2018-03-26 |
| Paper # | CCS2017-33 |
| Volume (vol) | vol.117 |
| Number (no) | CCS-520 |
| Page | pp.pp.1-6(CCS), |
| #Pages | 6 |
| Date of Issue | 2018-03-19 (CCS) |