| Presentation | 2021-11-18 Memory Efficient Training of Neural ODE by Symplectic Adjoint Method Takashi Matsubara, Yuto Miyatake, Takaharu Yaguchi, |
|---|---|
| PDF Download Page |  PDF download Page Link |
| Abstract(in Japanese) | (See Japanese page) |
| Abstract(in English) | Neural ODE learns an ordinary differential equation using neural networks, thereby modeling a continuous-time dynamics and probabilistic model with high accuracy. However, it uses the same neural network repeatedly, the backpropagation algorithm for training consumes extraordinary memory. Hence, the adjoint method has been often used, which calculates a gradient by a numerical integration. It requires much computational cost to suppress the numerical errors. Sanz-Serna, 2016, proposed to use a symplectic integrator for the adjoint method, thereby obtaining the exact gradient. In this study, we combine this method with a checkpointing scheme, and balance the fast calculation and small memory requirements. |
| Keyword(in Japanese) | (See Japanese page) |
| Keyword(in English) | Neural ODE / ordinary differential equation / adjoint method / symplectic integrator |
| Paper # | CCS2021-23 |
| Date of Issue | 2021-11-11 (CCS) |
| Conference Information | |
| Committee | CCS |
|---|---|
| Conference Date | 2021/11/18(2days) |
| Place (in Japanese) | (See Japanese page) |
| Place (in English) | Osaka Univ. |
| Topics (in Japanese) | (See Japanese page) |
| Topics (in English) | |
| Chair | Tetsuya Asai(Hokkaido Univ.) |
| Vice Chair | Megumi Akai(Hokkaido Univ.) / Masaki Aida(TMU) |
| Secretary | Megumi Akai(TDK) / Masaki Aida(Mie Univ.) |
| 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 Complex Communication Sciences |
|---|---|
| Language | JPN |
| Title (in Japanese) | (See Japanese page) |
| Sub Title (in Japanese) | (See Japanese page) |
| Title (in English) | Memory Efficient Training of Neural ODE by Symplectic Adjoint Method |
| Sub Title (in English) | |
| Keyword(1) | Neural ODE |
| Keyword(2) | ordinary differential equation |
| Keyword(3) | adjoint method |
| Keyword(4) | symplectic integrator |
| 1st Author's Name | Takashi Matsubara |
| 1st Author's Affiliation | Osaka University(Osaka Univ.) |
| 2nd Author's Name | Yuto Miyatake |
| 2nd Author's Affiliation | Osaka University(Osaka Univ.) |
| 3rd Author's Name | Takaharu Yaguchi |
| 3rd Author's Affiliation | Kobe University(Kobe Univ.) |
| Date | 2021-11-18 |
| Paper # | CCS2021-23 |
| Volume (vol) | vol.121 |
| Number (no) | CCS-253 |
| Page | pp.pp.31-36(CCS), |
| #Pages | 6 |
| Date of Issue | 2021-11-11 (CCS) |