CC BY
11*
ALT'23
The 30th International Conference on Advanced Laser Technologies
P-I-14
Optical coprocessor and diffraction neural networks
N. Kazanskiy12, L. Doskolovich1'2, N. Ivliev1'2, A. Nikonorov12, V. Podlipnov12, V. Protzenko12, R. Skidanov12, V. Soifer12, D. Soshnikov12
1-IPSI RAS-Branch of the FSRC "Crystallography and Photonics" RAS, 443001 Samara, Russia 2- Samara National Research University, 443086 Samara, Russia
email: [email protected]
The exponential increase in the number of publications on Scopus data indicates the growing interest of scientists from different countries in the subject of optical computing [1]. The authors of the report present optical circuits with light modulators that implement matrix calculations, Fourier correlator circuits, including a volumetric holographic filter that provides parallel recognition of 7,500 objects. The presented results of designing one-dimensional photonic crystal resonators based on comb waveguides for differentiating and integrating optical signals justify the potentially achievable parameters of the corresponding photonic crystal devices. The calculated photonic crystal integrators and differentiators are tens of times more compact than the known solutions, the proposed devices are easily assembled into cascades, integrated on a crystal and interfaced with electronic components [2]. The authors of the report proposed a scheme of an optical coprocessor designed for analyzing video streams and implemented as a computer board (module) with which the computer exchanges data via a fast PCIe bus. The coprocessor implements a Fourier correlator circuit with an amplitude spatial light modulator at the input and a phase spatial modulator in the frequency plane. In this scheme, the external video stream coming to the camera is converted by an amplitude modulator into a coherent video stream, and then it is processed in a Fourier correlator with a phase spatial modulator in the frequency plane, which sets the mathematical core of processing. The transmission function of the modulator is determined by the phase function matrix transmitted to it from the computer. This allows you to perform exactly the processing that is necessary for a given video stream at a given time. An analysis of the progress in the performance of existing modulators and cameras that allow working with large data arrays shows that by 2025, the calculation speed with using the proposed system of 1.00 * 1019 bits/s is possible. The system turns out to be quite compact (140*80 * 80mm) and relatively low-energy (no more than 100W - an order of magnitude smaller than a high-performance graphics card). The report shows experimental results of the selection of image contours using a created mock-up sample of such a coprocessor using a phase function implementing the Laplace operator on the modulator. Optical neural networks can also be implemented on such a system, while the phase function of the diffraction optical element (DOE) implementing the neural network layer is reproduced on the modulator. The report presents the results of a computational experiment for recognizing digitals and a number of symbols using one or more DOE. The authors show that with certain physical parameters, using only one DOE, it is possible to achieve recognition of handwritten digits with a probability higher than 0.91.
[1] N.L. Kazanskiy, M.A. Butt, S.N. Khonina, Optical Computing: Status and Perspectives, Nanomaterials, vol. 12, 2171 (2022).
[2] P. Serafimovich, N. Kazanskiy, Photonics Elements for Sensing and Optical Conversions (CRC Press), Chapter 2 Photonic crystal cavities in integrated on-chip optical signal processing components, (2023).
