2022
DOI: 10.1002/lpor.202200213
|View full text |Cite
|
Sign up to set email alerts
|

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, ther… Show more

Help me understand this report

Search citation statements

Order By: Relevance

Paper Sections

Select...
5

Citation Types

0
11
0

Year Published

2023
2023
2024
2024

Publication Types

Select...
6
4

Relationship

0
10

Authors

Journals

citations
Cited by 21 publications
(11 citation statements)
references
References 37 publications
0
11
0
Order By: Relevance
“…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).…”
Section: Introductionmentioning
confidence: 97%
“…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).…”
Section: Introductionmentioning
confidence: 97%
“…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.…”
Section: Introductionmentioning
confidence: 99%
“…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.…”
Section: Introductionmentioning
confidence: 99%