Numerical simulations on optoelectronic deep neural network hardware based on self-referential holography

• Published in: Optical review (Tokyo, Japan) Vol. 30; no. 3; pp. 387 - 396
• Main Authors: Tomioka, Rio, Takabayashi, Masanori
• Format: Journal Article
• Language: English
• Published: Tokyo Springer Japan 01.06.2023
• Subjects: Atomic; International Symposium on Imaging; Japan; Lasers; Microwaves; Molecular; Optical and Plasma Physics; Optical Devices; Optical Memory (ISOM’ 22); Optics; Photonics; Physics; Physics and Astronomy; Quantum Optics; RF and Optical Engineering; Sapporo; Sensing; Special Section: Regular Paper; Optical neural network; Self-referential holography; Optical information processing; Holographic data storage
• ISSN: 1340-6000; 1349-9432
• DOI: 10.1007/s10043-023-00810-2

We propose a novel optoelectronic deep neural network (OE-DNN) hardware called the self-referential holographic deep neural network (SR-HDNN). The SR-HDNN features a combination of an optical computing part utilizing a volume hologram and an electronic part connecting the optical elements virtually. Since the shape of a volume hologram, which is a 3-dimensional (3D) refractive index distribution in this case, can be changed by its recording conditions, it is expected to realize the flexible design of optical computing functions by coupling between specific nodes. In addition, the electronic part enables the construction of multi-layer networks without extending the optical system and enabling arbitrary signal processing, including nonlinear operations. By integrating flexible optical and electronic parts, the SR-HDNN consisting of both flexible optical and electronic parts has the potential to maximize the performance of OE-DNN. In this study, we numerically simulate image classification tasks to investigate the feasibility and potential of the SR-HDNN.