| Presentation | 2021-08-05 Digital Implement of 3-layered Neural Networks with Stochastic Activation, Shunting Inhibition, and a Dual-rail Backpropagation Yoshiaki Sasaki, Seiya Muramatsu, Kohei Nishida, Megumi Akai-Kasaya, Tetsuya Asai, |
|---|---|
| PDF Download Page |  PDF download Page Link |
| Abstract(in Japanese) | (See Japanese page) |
| Abstract(in English) | Stochastic computing (SC) is an arithmetic technique that enables various operations to be performed with a small number of logic gates in exchange for operational accuracy. In this paper, we discuss the realization of an edge AI integrated circuit with a learning function based on the SC. First, we propose a logic circuit that realizes an area-efficient activation function that does not require decoding in the midst of forward operations. Next, we propose a method to realize an operation equivalent to subtraction, which is considered to be difficult in SC, as well as propose a new arithmetic method that mimics the synaptic transmission of neural circuits. By assuming the input to the activation function, we also propose a summation operation that utilizes imperfect addition, which has rarely been used in SC. By introducing these arithmetic methods, we propose a digital implementation of a multilayer perceptron based on the SC. First, we learn simple logic operations and then learn linear and nonlinear regression problems. We then demonstrate the feasibility of SC-based machine learning. The method presented herein will expand the options of edge AI integrated circuits using SC and may contribute to the development of edge AI integrated circuits in the future. |
| Keyword(in Japanese) | (See Japanese page) |
| Keyword(in English) | Stochastic Computing / Machine Learning / Neural Network / Edge AI Integrated Circuit / Backpropagation |
| Paper # | CCS2021-16 |
| Date of Issue | 2021-07-29 (CCS) |
| Conference Information | |
| Committee | IN / CCS |
|---|---|
| Conference Date | 2021/8/5(1days) |
| Place (in Japanese) | (See Japanese page) |
| Place (in English) | Online |
| Topics (in Japanese) | (See Japanese page) |
| Topics (in English) | Network Science, Future Network, Cloud/SDN/Virtualization, Contents Delivery/Contents Exchange, and others |
| Chair | Kenji Ishida(Hiroshima City Univ.) / Tetsuya Asai(Hokkaido Univ.) |
| Vice Chair | Kunio Hato(Internet Multifeed) / Megumi Akai(Hokkaido Univ.) / Masaki Aida(TMU) |
| Secretary | Kunio Hato(NTT) / Megumi Akai(Univ. of Nagasaki) / Masaki Aida(Nagaoka Univ. of Tech.) |
| Assistant | / Hidehiro Nakano(Tokyo City Univ.) / Hiroyasu Ando(Tsukuba Univ.) / Takashi Matsubara(Osaka Univ.) / Sumiko Miyata(Shibaura Inst. of Tech.) |
| Paper Information | |
| Registration To | Technical Committee on Information Networks / Technical Committee on Complex Communication Sciences |
|---|---|
| Language | JPN |
| Title (in Japanese) | (See Japanese page) |
| Sub Title (in Japanese) | (See Japanese page) |
| Title (in English) | Digital Implement of 3-layered Neural Networks with Stochastic Activation, Shunting Inhibition, and a Dual-rail Backpropagation |
| Sub Title (in English) | |
| Keyword(1) | Stochastic Computing |
| Keyword(2) | Machine Learning |
| Keyword(3) | Neural Network |
| Keyword(4) | Edge AI Integrated Circuit |
| Keyword(5) | Backpropagation |
| 1st Author's Name | Yoshiaki Sasaki |
| 1st Author's Affiliation | Hokkaido University(Hokkaido Univ.) |
| 2nd Author's Name | Seiya Muramatsu |
| 2nd Author's Affiliation | Hokkaido University(Hokkaido Univ.) |
| 3rd Author's Name | Kohei Nishida |
| 3rd Author's Affiliation | Hokkaido University(Hokkaido Univ.) |
| 4th Author's Name | Megumi Akai-Kasaya |
| 4th Author's Affiliation | Hokkaido University(Hokkaido Univ.) |
| 5th Author's Name | Tetsuya Asai |
| 5th Author's Affiliation | Hokkaido University(Hokkaido Univ.) |
| Date | 2021-08-05 |
| Paper # | CCS2021-16 |
| Volume (vol) | vol.121 |
| Number (no) | CCS-134 |
| Page | pp.pp.7-13(CCS), |
| #Pages | 7 |
| Date of Issue | 2021-07-29 (CCS) |