Efficient stochastic parallel gradient descent training for on-chip optical processors

A new publication from Opto-Electronic Advances; DOI  10.29026/oea.2024.230182 , discusses efficient stochastic parallel gradient descent training for on-chip optical processors.

 

With the explosive growth of global data volume, space-division multiplexing (SDM) technology has been emerged as a promising solution to enhance the communication capacity. Over the past few decades, SDM has been realized in few-mode fibers, multi-core fiber and free-space optical communication systems. However, all of above solutions face challenges of signal crosstalk because of the mixing between different channels during the transmission of optical signals, resulting in a degradation in signal quality at the receiver. Therefore, digital signal processing (DSP) is necessitated for descrambling. Unfortunately, high-speed DSP chips in the electrical domain are highly complicated, difficult to design, and high power consumption. In recent years, integrated reconfigurable optical processors have been exploited to undo the channel mixing in the optical domain. However, the gradient descent algorithms need to update variables one by one to calculate the loss function with each iteration, which leads to a large amount of computation and a long training time. What’s more, the swarm intelligence algorithms such as genetic algorithm (GA) and particle swarm optimization (PSO) algorithm have a large enough population size to ensure the reliability of the training results, which also bring a large amount of computation. Therefore, it is of great significance to find an efficient optimization algorithm suitable for optical matrix configuration for online training of large-scale photonic computing chips and multi-dimensional optical communication systems.

 

Progress in the online training of optical matrix computing chips has been made; compared to the discrete gradient descent algorithm, GA and PSO algorithm, this method greatly reduces the number of operations, which can greatly save the power consumption in the training process, and is expected to be applied to the online training of ultra-large-scale optical matrix computing chips.

 

In order to verify the feasibility of the proposed optimization method, the authors designed and fabricated a 6×6 reconfigurable optical processor chip based on cascaded Mach-Zehnder Interferometers (MZIs) and carried out online training experiments, including optical switching matrix and optical signal descrambling matrix. Figure 2 shows the application scenario of an optical processor in the MDM optical communication system and the internal structure of the processor. Figure 2 shows the training results, and it can be seen that the training effect is relatively good for the optical switching and optical signal descrambling tasks in the multi-dimensional optical communication system.

 

On this basis, this reconfigurable optical processor chip to high-speed optical communication systems was used to compensate for crosstalk caused by mode mixing during transmission. Figure 3 shows the experimental setup and obtained results. It can be seen that the quality of the signal is significantly improved when the light passes through the trained photonic chip, as shown in Figs. 3(e) and 3(f), and the bit error rate (BER) is greatly reduced.

 

Finally, the computational effort of the SPGD algorithm was compared with the traditional gradient algorithm, GA and PSO algorithm when the optical matrix scale is expanded to 10×10, 16×16, 32×32. The results show that the increase of the computational cost of SPGD algorithm is less than that of other algorithms.

 

Table 1 Comparison of different algorithms

algorithm	numbers of update	matrix sizes
		N=6	N=10	N=16	N=32
GD	N(N-1)×T	690	3870	13200	93248
GA	M×T	1048	9046.67	39732	171200
PSO	M×T	1024	5912	31056	116145
SPGD	3×T	297.9	1092.6	4752.6	18053.1

 

Keywords: optical communications / optical signal processing / channel descrambling / optical neural network chip / silicon photonics

# # # # # #

The Multi-Dimensional Photonics Laboratory (MDPL) at Huazhong University of Science and Technology has been dedicated to the research in light field manipulation, multi-dimensional optical communications, optical signal processing, optoelectronic devices and integration, and silicon-based photonic integrated circuits, and has made a series of important achievements. The team targets national major demands and world frontiers in science and technology, such as high-speed large-capacity optical communications and photonic integrated circuits and undertakes multiple National Key R&D Program of China and other projects. The related achievements have been selected for the important progress of the Optical Society and the National 13th Five Year Plan Science and Technology Innovation Achievement Exhibition.

 

Professor Wang Jian, the team leader of MDPL, is the vice director of Wuhan National Laboratory for Optoelectronics (WNLO). He is a recipient of the National Distinguished Young Science Foundation supported by the National Natural Science Foundation of China (NSFC). He has been elected as a Fellow of the Institute of Electrical and Electronics Engineers (IEEE Fellow), a fellow of the Optical Society (OPTICA Fellow), and a fellow of the International Society for Optical Engineering (SPIE Fellow). He is also a recipient of the National Outstanding Young Science Foundation, Young Changjiang Scholars of China, National Youth Top-notch Talent of China, and Royal Society-Newton Advanced Fellowship. He serves as the OPTICA Fellow Members Committee, executive director of the Chinese Optical Society, and vice chair of the IEEE Photonics Society Wuhan Chapter. He has been selected as one of the top 2% global scientists from 2019 to 2023 and a highly cited Chinese scholar by Elsevier from 2020 to 2022. As the first author, he has won two first prize in Natural Science Award of Ministry of Education of China, one Youth Science Award of Ministry of Education of China, one first prize in Natural Science of the Chinese Optical Society, and one Wang Daheng Optics Award for Young and Middle-aged Scientific and Technological Personnel.

# # # # # #

Opto-Electronic Advances (OEA) is a rapidly growing high-impact, open access, peer reviewed monthly SCI journal with an impact factor of 14.1 (Journal Citation Reports for IF2022). Since its launch in March 2018, OEA has been indexed in SCI, EI, DOAJ, Scopus, CA and ICI databases over the time, and expanded its Editorial Board to 30 members from 17 countries.

The journal is published by The Institute of Optics and Electronics, Chinese Academy of Sciences, aiming at providing a platform for researchers, academicians, professionals, practitioners, and students to impart and share knowledge in the form of high quality empirical and theoretical research papers covering the topics of optics, photonics and optoelectronics.

# # # # # #

 

More information: http://www.oejournal.org/oea

Editorial Board: http://www.oejournal.org/oea/editorialboard/list

All issues available in the online archive (http://www.oejournal.org/oea/archive).

Submissions to OEA may be made using ScholarOne (https://mc03.manuscriptcentral.com/oea).

ISSN: 2096-4579

CN: 51-1781/TN

Contact Us: oea@ioe.ac.cn

Twitter: @OptoElectronAdv (https://twitter.com/OptoElectronAdv?lang=en)

WeChat: OE_Journal

# # # # # #

 

Wan YJ, Liu XD, Wu GZ et al. Efficient stochastic parallel gradient descent training for on-chip optical processor. Opto-Electron Adv 7, 230182 (2024). doi: 10.29026/oea.2024.230182 

 

---

---