High‐Throughput Multichannel Parallelized Diffraction Convolutional Neural Network Accelerator

Abstract

Convolutional neural networks are paramount in image and signal processing, and are responsible for the majority of image recognition power consumption today, concentrated mainly in convolution computations. With convolution operations being computationally intensive, next‐generation hardware accelerators need to offer parallelization and high efficiency. Diffractive optics offers the promise of low‐latency, highly parallel convolution operations. However, thus far parallelism is only partially harvested, thereby significantly underdelivering in comparison to its throughput potential. Here, a parallelized operation high‐throughput Fourier optic convolutional accelerator is demonstrated. For the first time, simultaneous processing of multiple kernels in Fourier domain enabled by optical diffraction orders is achieved alongside input parallelism. The proposed approach can offer ≈100× speedup over the previous generation optical diffraction‐based processor and 10× speedup over other state‐of‐the‐art optical Fourier classifiers.

Free-space optical multiplexed orbital angular momentum beam identification system using Fourier optical convolutional layer based on 4f system

“…After two decades development, kernels generated by state-in-the-art DMD outperforms the older hologram mask with a higher enough throughput to simulate completely neural planes, makes optical convolutional efficiency promising. [87][88][89][90][91][92][93][94][95][96][97][98][99][100][101] In this study, we experimentally generate multiplexed OAM beams and introduce optical filtering method which relys on spatial Fourier transform of images in the frequency domain as optical convolutional neural network to train and identify multiplexed OAM beams under simulated atmospheric turbulence conditions, we show the system currently capable of classifying 12 classes at test accuracy of 95% (under weak turbulence) and 87% (under strong turbulence).…”

High-performance optoelectronics for integrated photonic neural networks

“…Artificial neurons based on nanophotonic technologies can potentially provide the platform that can fulfill the challenging future technological needs. [1][2][3][4][5][6][7][8][9][10][11][12][13][14][15][16][17][18][19][20] Integrated photonic technology provides a solution to the limitations of current digital electronic counterparts like efficient fundamental computational operations such as weighted sum or addition, vector matrix multiplications, or convolutions technologically enabled by attojoule efficient electro-optic (EO) modulators, phase shifters, and combiners. Furthermore, high parallelism and bandwidth is provided by exploiting wavelength-, polarization-and/or mode-division multiplexing.…”

Michelson interferometric methods for full optical complex convolution

“…Optical real-time dynamic data processing has several applications including accelerating tensor algebra, which usually creates a bottle-neck in in silico machine learning, 1-10 cryptography, [11][12][13][14][15][16][17] digital holography, [18][19][20][21][22][23][24] and more optical computing algorithms [25][26][27][28][29][30][31][32][33][34] is given their high computational overhead. Alongside that, cryptography algorithms are highly dependent on computationally hungry data compression algorithms to provide the necessary level of security and meet common network specifications, while image processing Machine Learning algorithms dependent heavily on the speed of convolution and area dot product multiplication as their major computation budget consumer.…”