Hybrid Convolutional Optoelectronic Reservoir Computing for Image Recognition

Philip Jacobson, Mizuki Shirao, Kerry Yu, Guan-Lin Su and Ming C. Wu

EECS Department
University of California, Berkeley
Technical Report No. UCB/EECS-2021-251
December 8, 2021

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2021/EECS-2021-251.pdf

Photonic delay-based reservoir computers (RC) have emerged as an attractive high-speed, low-power alternative to traditional digital hardware for AI. We demonstrate experimentally a novel hybrid RC scheme in which input data is first preprocessed through several convolutional layers, either trained or untrained, digitally to generate novel feature maps. These random feature maps are then processed through an optoelectronic implementation of delay-based RC. Using the MNIST dataset of handwritten digits, experiments of our proposed hybrid scheme achieve classification error of 1.6% using untrained convolutions, and an error of 1.1% using trained convolutions, results comparable to that of state-of-the-art machine learning algorithms. Additionally, our experimental implementation can offer a potential 10 decrease in model training time, compared to that of common digital alternatives.

Advisor: Ming C. Wu


BibTeX citation:

@mastersthesis{Jacobson:EECS-2021-251,
    Author = {Jacobson, Philip and Shirao, Mizuki and Yu, Kerry and Su, Guan-Lin and Wu, Ming C.},
    Title = {Hybrid Convolutional Optoelectronic Reservoir Computing for Image Recognition},
    School = {EECS Department, University of California, Berkeley},
    Year = {2021},
    Month = {Dec},
    URL = {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2021/EECS-2021-251.html},
    Number = {UCB/EECS-2021-251},
    Abstract = {Photonic delay-based reservoir computers (RC) have
emerged as an attractive high-speed, low-power alternative to
traditional digital hardware for AI. We demonstrate experimentally a novel hybrid RC scheme in which input data is first preprocessed through several convolutional layers, either trained or untrained, digitally to generate novel feature maps. These random feature maps are then processed through an optoelectronic implementation of delay-based RC. Using the MNIST dataset of handwritten digits, experiments of our proposed hybrid scheme achieve classification error of 1.6% using untrained convolutions, and an error of 1.1% using trained convolutions, results comparable to that of state-of-the-art machine learning algorithms. Additionally, our experimental implementation can offer a potential 10 decrease in model training time, compared to that of common digital alternatives.}
}

EndNote citation:

%0 Thesis
%A Jacobson, Philip
%A Shirao, Mizuki
%A Yu, Kerry
%A Su, Guan-Lin
%A Wu, Ming C.
%T Hybrid Convolutional Optoelectronic Reservoir Computing for Image Recognition
%I EECS Department, University of California, Berkeley
%D 2021
%8 December 8
%@ UCB/EECS-2021-251
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2021/EECS-2021-251.html
%F Jacobson:EECS-2021-251