| Presentation | 2020-01-31 [Poster Presentation] Communication-Efficient Federated Learning Using Non-Labeled Data Souhei Itahara, Takayuki Nishio, Masahiro Morikura, Koji Yamamoto, |
|---|---|
| PDF Download Page |  PDF download Page Link |
| Abstract(in Japanese) | (See Japanese page) |
| Abstract(in English) | Federated learning (FL) is a machine learning setting where many mobile devices collaboratively train a machine learning (ML) model, while keeping the training data decentralized. In FL, each device updates a model with his/her data and uploads the model to a server which aggregates the models instead of uploading the training data to the server.Thus, the traffic for uploading the model is not negligible.This paper proposes a cooperative learning method, called Distillation Based Semi-Supervised Federated Learning (DS-FL), which aims to reduce traffic required for training the ML model. In DS-FL, non-labeled open data is used for the cooperative model training via semi-supervised learning.Each device trains a model with his/her data, predicts logits for the open data, and updates the model with the open data and aggregated logits. Since the data size of the logits is much smaller than that of the models, traffic is reduced largely. We evaluate our method using an image classification task (MNIST). Our experiments show that the proposed method achieves 94% less traffic than that of the previous method. |
| Keyword(in Japanese) | (See Japanese page) |
| Keyword(in English) | Federated Learning / Semi-Supervised Learning / Machine Learning / Communication Cost |
| Paper # | SeMI2019-109 |
| Date of Issue | 2020-01-23 (SeMI) |
| Conference Information | |
| Committee | SeMI |
|---|---|
| Conference Date | 2020/1/30(2days) |
| Place (in Japanese) | (See Japanese page) |
| Place (in English) | |
| Topics (in Japanese) | (See Japanese page) |
| Topics (in English) | |
| Chair | Susumu Ishihara(Shizuoka Univ.) |
| Vice Chair | Kazuya Monden(Hitachi) / Koji Yamamoto(Kyoto Univ.) |
| Secretary | Kazuya Monden(Kyoto Univ.) / Koji Yamamoto(NTT DOCOMO) |
| Assistant | Akira Uchiyama(Osaka Univ.) / Kenji Kanai(Waseda Univ.) / Masafumi Hashimoto(Osaka Univ.) |
| Paper Information | |
| Registration To | Technical Committee on Sensor Network and Mobile Intelligence |
|---|---|
| Language | JPN |
| Title (in Japanese) | (See Japanese page) |
| Sub Title (in Japanese) | (See Japanese page) |
| Title (in English) | [Poster Presentation] Communication-Efficient Federated Learning Using Non-Labeled Data |
| Sub Title (in English) | |
| Keyword(1) | Federated Learning |
| Keyword(2) | Semi-Supervised Learning |
| Keyword(3) | Machine Learning |
| Keyword(4) | Communication Cost |
| 1st Author's Name | Souhei Itahara |
| 1st Author's Affiliation | Kyoto University(Kyoto Univ) |
| 2nd Author's Name | Takayuki Nishio |
| 2nd Author's Affiliation | Kyoto University(Kyoto Univ) |
| 3rd Author's Name | Masahiro Morikura |
| 3rd Author's Affiliation | Kyoto University(Kyoto Univ) |
| 4th Author's Name | Koji Yamamoto |
| 4th Author's Affiliation | Kyoto University(Kyoto Univ) |
| Date | 2020-01-31 |
| Paper # | SeMI2019-109 |
| Volume (vol) | vol.119 |
| Number (no) | SeMI-406 |
| Page | pp.pp.47-48(SeMI), |
| #Pages | 2 |
| Date of Issue | 2020-01-23 (SeMI) |