Optimization of optical convolution kernel of optoelectronic hybrid convolution neural network

Abstract

To enhance the optical computation’s utilization efficiency, we develop an optimization method for optical convolution kernel in the optoelectronic hybrid convolution neural network (OHCNN). To comply with the actual calculation process, the convolution kernel is expanded from single-channel to two-channel, containing positive and negative weights. The Fashion-MNIST dataset is used to test the network architecture’s accuracy, and the accuracy is improved by 7.5% with the optimized optical convolution kernel. The energy efficiency ratio (EER) of two-channel network is 46.7% higher than that of the single-channel network, and it is 2.53 times of that of traditional electronic products.

References

[1] LiS, SunK, LuoY, et al.. Novel CNN-based AP2D-net accelerator: an area and power efficient solution for real-time applications on mobile FPGA[J]. Electronics, 2020, 9(5):832

[2] WuY, Abdel-AtyM, ZhengO, et al.. Automated safety diagnosis based on unmanned aerial vehicle video and deep learning algorithm[J]. Transportation research record: journal of the transportation research board, 2020, 2674(99): 036119812092580

[3] PengH, NahmiasM A, ThomasF, et al.. Neuromorphic photonic integrated circuits[J]. IEEE journal of selected topics in quantum electronics, 2018, 24(6): 1-15

[4] ShenY, HarrisN C, SkirloS, et al.. Deep learning with coherent nanophotonic circuits[J]. Nature photonics, 2017, 11: 441-446

[5] BangariV, MarquezB A, MillerH, et al.. Digital electronics and analog photonics for convolutional neural networks (DEAP-CNNs)[J]. IEEE journal of selected topics in quantum electronics, 2020, 26(1):1-13

[6] OngJ R, OoiC C, AngT Y L, et al.. Photonic convolutional neural networks using integrated diffractive optics[J]. IEEE journal of selected topics in quantum electronics, 2020, 26(5):1-8

[7] PaiS, BartlettB, SolgaardO, et al.. Matrix optimization on universal unitary photonic devices[J]. Physical review applied, 2019, 11: 064044

[8] YingZ F, FengC H, ZhaoZ, et al.. Electronic-photonic arithmetic logic unit for high-speed computing[J]. Nature communications, 2020, 11: 2154

[9] XuX, ZhuL, ZhuangW, et al.. Photoelectric hybrid convolution neural network with coherent nanophotonic circuits[J]. Optical engineering, 2021, 60(11):117106

[10] ClementsW R, HumphreysP C, MetcalfB J, et al.. An optimal design for universal multiport interferometers[J]. Optica, 2016, 3(12): 1460

[11] ReckM, ZeilingerA, BernsteinH J, et al.. Experimental realization of any discrete unitary operator[J]. Physical review letters, 1994, 73(1): 58-61

[12] ConnellyM J. Semiconductor optical amplifiers[M], 2007, Berlin, Heidelberg, Springer Science & Business Media

[13] CORPORATION N. NVIDIA TESLA V100[EB/OL]. (2018-01-26) [2021-12-04]. https://www.nvidia.com/en-us/data-center/tesla-v100/.

[14] MillerD A B. Device requirements for optical interconnects to silicon chips[J]. Proceedings of the IEEE, 2009, 97(7):1166-1185

[15] FeldmannJ, YoungbloodN, KarpovM, et al.. Parallel convolutional processing using an integrated photonic tensor core[J]. Nature, 2021, 589: 52-58

[16] JhaA, HuangC, PrucnalP R. Reconfigurable all-optical nonlinear activation functions for neuromorphic photonics[J]. Optics letters, 2020, 45(17):4819-4822