| IEICE Technical Committee Submission System Conference Paper's Information |
Online Proceedings [Sign in] Tech. Rep. Archives |
| Go Top Page | Go Previous | [Japanese] / [English] |
| Paper Abstract and Keywords | ||
| Presentation | 2024-06-06 09:55 Training approaches for photonic reservoir computing circuits Keigo Takabayashi, Ryota Nakayama, Takeo Maruyama, Tomoaki Niiyama, Satoshi Sunada (Kanazawa Univ.)  NLP2024-15 CCS2024-2 |
|
| Abstract | (in Japanese) | (See Japanese page) |
| (in English) | The rapid development of information technology in recent years has led to an explosive increase in the demand for computing for advanced and diverse tasks. Optical reservoir computing (optical RC), which enables high-speed information processing with low power consumption, has attracted much attention in recent years[1],[2]. Optical RC is a method for performing complex computational tasks using a nonlinear dynamic system (reservoir), and is particularly effective for forecasting time-series data[3]. Although RC with memory functions are important for time series data processing, on-chip RC have the problem of not being able to reserve a large memory area. In this study, we proposed and fabricated an optical reservoir circuit with a memory retention function by repeatedly combining and branching multiple waveguides in an interferometer to create complex interference, and evaluated its basic performance. In addition, we implemented an online learning system in which all inference is performed by light by coupling a readout layer with 22 electrodes using a Mach–Zehnder interference optical switch (MZI) to the back-end of the optical reservoir, and conducted a simple evaluation of this system. The experimental system is shown in Fig. 1. A laser beam of 1550nm wavelength is phase-modulated into a time-series signal at a rate of 25 GS/s and sent directly to an optical RC chip as an input. This RC chip is fabricated on a silicon substrate and, as described above, obtains the reservoir response by multiple interference of light passing through multiple branching waveguides and outputs it from eight output ports. We propose a system that approaches the target waveform by performing optimization processing using CMA-ES in the readout layer with MZI coupled to the output port. In this experiment, we evaluated the time series prediction performance using chaotic waveforms and the memory retention performance using random signals. For the evaluation of the optical reservoir part, it was shown that the system can predict one step ahead (Fig. 2(a)) of the chaotic waveform with NMSE=0.1 or less. It was also found that prediction up to 20 steps ahead (Fig. 2(b)) was also possible with a low NMSE. Furthermore, in a system where the entire inference is performed optically, we showed that it is possible to predict a chaotic waveform one step ahead with an NMSE ≈ 0.3, indicating the possibility of implementing an all-optical readout layer with MZI, which enables low latency and high speed processing. |
|
| Keyword | (in Japanese) | (See Japanese page) |
| (in English) | optical circuit / mach-zehnder optical interferometry / neural network / covariance matrix adaptive evolution strategy / / / / | |
| Reference Info. | IEICE Tech. Rep., vol. 124, no. 63, CCS2024-2, pp. 7-7, June 2024. | |
| Paper # | CCS2024-2 | |
| Date of Issue | 2024-05-30 (NLP, CCS) | |
| ISSN | Online edition: ISSN 2432-6380 | |
| Copyright and reproduction |
All rights are reserved and no part of this publication may be reproduced or transmitted in any form or by any means, electronic or mechanical, including photocopy, recording, or any information storage and retrieval system, without permission in writing from the publisher. Notwithstanding, instructors are permitted to photocopy isolated articles for noncommercial classroom use without fee. (License No.: 10GA0019/12GB0052/13GB0056/17GB0034/18GB0034) | |
| Download PDF | NLP2024-15 CCS2024-2 |
|
| Conference Information | |
| Committee | NLP CCS |
| Conference Date | 2024-06-06 - 2024-06-07 |
| Place (in Japanese) | (See Japanese page) |
| Place (in English) | West Japan General Exhibition Center AIM |
| Topics (in Japanese) | (See Japanese page) |
| Topics (in English) | Nonlinear Problems, Complex Communication Sciences, etc. |
| Paper Information | |
| Registration To | CCS |
| Conference Code | 2024-06-NLP-CCS |
| Language | Japanese |
| Title (in Japanese) | (See Japanese page) |
| Sub Title (in Japanese) | (See Japanese page) |
| Title (in English) | Training approaches for photonic reservoir computing circuits |
| Sub Title (in English) | |
| Keyword(1) | optical circuit |
| Keyword(2) | mach-zehnder optical interferometry |
| Keyword(3) | neural network |
| Keyword(4) | covariance matrix adaptive evolution strategy |
| Keyword(5) | |
| Keyword(6) | |
| Keyword(7) | |
| Keyword(8) | |
| 1st Author's Name | Keigo Takabayashi |
| 1st Author's Affiliation | Kanazawa University (Kanazawa Univ.) |
| 2nd Author's Name | Ryota Nakayama |
| 2nd Author's Affiliation | Kanazawa University (Kanazawa Univ.) |
| 3rd Author's Name | Takeo Maruyama |
| 3rd Author's Affiliation | Kanazawa University (Kanazawa Univ.) |
| 4th Author's Name | Tomoaki Niiyama |
| 4th Author's Affiliation | Kanazawa University (Kanazawa Univ.) |
| 5th Author's Name | Satoshi Sunada |
| 5th Author's Affiliation | Kanazawa University (Kanazawa Univ.) |
| Speaker | Author-1 |
| Date Time | 2024-06-06 09:55:00 |
| Presentation Time | 25 minutes |
| Registration for | CCS |
| Paper # | NLP2024-15, CCS2024-2 |
| Volume (vol) | vol.124 |
| Number (no) | no.62(NLP), no.63(CCS) |
| Page | p.7 |
| #Pages | 1 |
| Date of Issue | 2024-05-30 (NLP, CCS) |
[Return to Top Page]