Demultiplexing OAM beams via Fourier optical convolutional neural network

Abstract

Here we present an innovative free-space optical (FSO) communication system which is capable of training database in real-time and demultiplex multiplexed spatial structured laser beams such as orbital angular momentum (OAM) beams under varying atmospheric turbulent conditions. The core part of our detection system is heterogeneous convolutional neural network includes an optical 4f system using first Fourier convolution neural network layer driven by kilohertz-fast reprogrammable high-resolution digital micromirror devices (DMDs). This optical-filtering-based convolutional neural network is utilized to realize the training and demultiplexing 4-bit OAM-coded beams under simulated turbulent condition using modified von K´arm´an atmospheric model. The current implementation shows classification accuracy of 89.35% (under weak turbulence) and 38.26% (under strong turbulence).

An efficient optical-based binary neural network hardware accelerator for harsh environments

“…This unprecedented growth is not merely a quantitative leap but signifies a qualitative transformation in computational capabilities, enabling nuanced understanding and processing of information at scales hitherto unattainable. [3][4][5][6][7] As we navigate through this era of burgeoning model sizes, it becomes imperative to scrutinize the implications of such advancements. The integration of vast parameters necessitates a reevaluation of computational efficiency, the sustainability of energy consumption, and the ethical considerations surrounding the deployment of these models.…”

Low-latency full precision optical convolutional neural network accelerator

“…6 The rapid escalation in the computational demands of CNNs has significantly outstripped the predictions of Moore's Law, especially over the past decade. [8][9][10] Fig. 1(a) shows the exponential CNN parameter increase, indicating a need for solutions beyond traditional hardware accelerators.…”