Programmable low-threshold optical nonlinear activation functions for photonic neural networks

Chu, Tao; Huang, Ying Sophie; Wei-ping, Wang; Qiao, Lei; Hu, Xianliang

doi:10.1364/ol.451287

Cited by 19 publications

(14 citation statements)

References 14 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Reference [55] proposed two programmable all-optical non-linear activation methods, which are implemented on silicon-on-insulator (SOI) substrates. The devices include Ge-Si PIN photodetectors (PDs) and the PIN electro-optic (EO) F I G U R E 9 Optical implementation of the 4 � 4 matrix using MZI [49] F I G U R E 1 0 Schematic diagram of the activation function circuit for generating non-linear function f (z) [54] DU ET AL.…”

Section: All-optical Methodsmentioning

confidence: 99%

“…Reference [55] proposed two programmable all‐optical non‐linear activation methods, which are implemented on silicon‐on‐insulator (SOI) substrates. The devices include Ge‐Si PIN photodetectors (PDs) and the PIN electro‐optic (EO) switch.…”

Section: Non‐linear Activation Functionmentioning

confidence: 99%

“…The non‐linear activation is realised through the above steps. In reference [55], the experimental results show that the optical threshold power of the device is 0.2 MW and the measured non‐linear activation function is applied to the convolutional neural network for the MNIST handwritten numeral classification task. Using these functions, the accuracy of the convolutional neural network is up to 95.71%.…”

Section: Non‐linear Activation Functionmentioning

confidence: 99%

“… Device structure for realising all‐optical non‐linear activation proposed in reference [55] includes (a) MZI or (b) MRR EO switch and Ge‐Si PD. [Reprinted/Adapted] with permission from [ref #] © The Optical Society …”

Section: Non‐linear Activation Functionmentioning

confidence: 99%

See 3 more Smart Citations

Implementation of optical neural network based on Mach–Zehnder interferometer array

Yuan

et al. 2022

IET Optoelectronics

View full text Add to dashboard Cite

Compared with electrons, photons have the potential to realise ultra-high speed operations because of its unique high speed and high parallelism. In recent years, there have been many researches on neural networks using optical hardware. The Mach-Zehnder interferometer (MZI) and micro-ring resonator (MRR) are commonly used as optical devices to realise linear operation units in optical neural networks (ONN). MZI has the advantages of simple fabrication, high sensitivity, and easy integration, which has attracted the attention of researchers. We summarise the implementation methods of ONN matrix multiplication based on MZI, the implementation methods of non-linear activation, and the on-chip training methods. We first summarise the researches on matrix multiplication of ONN based on MZI. Three kinds of MZI grid decomposition methods, Fast Fourier Transform (FFT) grid structures, and the corresponding derivation processes are introduced, respectively. Then, several experimental implementations of ONN based on MZI are summarised, and the characteristics of optical processors fabricated in these references are analysed. Finally, the realisation methods of non-linear activation and on-chip training of silicon ONN are introduced, respectively.

show abstract

Section: All-optical Methodsmentioning

confidence: 99%

Section: Non‐linear Activation Functionmentioning

confidence: 99%

Section: Non‐linear Activation Functionmentioning

confidence: 99%

Section: Non‐linear Activation Functionmentioning

confidence: 99%

See 2 more Smart Citations

Implementation of optical neural network based on Mach–Zehnder interferometer array

Yuan

et al. 2022

IET Optoelectronics

View full text Add to dashboard Cite

show abstract

“…The linear operation can be realized by using the coherence and superposition of linear optics, [7] but nonlinear operations are usually performed electronically, which does not facilitate the realization of all-optical neural networks (AONNs). One of the current obstacles to the implementation of ONNs is the lack of optical non linearity, [8] because the complex computational requirements cannot be met by linear optical matrix operations alone. [9] In a neural network, a nonlinear activation function can accelerate the convergence of the network and improve the recognition accuracy, which is an indispensable component of the ONN.…”

mentioning

confidence: 99%

MXene‐Based Broadband Ultrafast Nonlinear Activator for Optical Computing

Yang

Tan

Zhang

et al. 2022

Advanced Optical Materials

View full text Add to dashboard Cite

Optical neural networks (ONNs) are particularly advantageous owing to their inherent parallelism and low energy consumption. However, one of the obstacles to the implementation of ONNs is the lack of optical nonlinearity. In this study, optical nonlinear activators for ONNs are prepared by combining Ti3C2Tx MXene with microfibers and their principles are verified. Activation functions obtained from experimental measurements are used to simulate multiclassification and super‐resolution reconstruction tasks with performance comparable to that of activation functions commonly used in computers. Four necessary criteria are proposed and validated for evaluating the performance of the nonlinear activator: recovery time, deviation from linearity, the activation function close to identity mapping, and reconfigurability of the configuration. Theoretically, the nonlinear activator can compute 100 times faster than commonly used electronic computers and can be used as a nonlinear activation unit for ONNs to help the integration of ONNs with artificial intelligence.

show abstract

Dual Optical Frequency Comb Neuron: Co‐Developing Hardware and Algorithm

Zhang

Tao

Yan

et al. 2023

Advanced Intelligent Systems

View full text Add to dashboard Cite

Previous studies on photonic neural network have demonstrated that algorithm can inspire hardware design. This study seeks to demonstrate that hardware can also inspire algorithm design. To further exploit the advantages of photonic analog computing, the authors develop hardware and algorithm simultaneously for photonic convolutional neural networks. Specifically, this work developed an architecture called dual optical frequency comb neuron (DOFCN) enabled by an integrated microcomb to perform cosinusoidal nonlinear activation and vector convolution without temporal or spatial dispersion and large‐scale modulator arrays. Furthermore, DOFCN‐based composite vector convolutional neural networks (CVCNNs), an optical‐electric hybrid model, are proposed to perform classification and regression tests in signal modulation format identification and optical structure inverse design tasks, respectively. The ablation experiments show that under 4‐bit precision limit, the element‐wise activation CVCNN has 14% higher classification accuracy, 76% lower regression residuals, and 100% higher training efficiency than that of the 32‐bit standard convolutional neural network (CNN). DOFCN exhibits impressive spectral information processing ability to facilitate signal‐processing tasks related to optics and electromagnetics.

show abstract

Programmable low-threshold optical nonlinear activation functions for photonic neural networks

Cited by 19 publications

References 14 publications

Implementation of optical neural network based on Mach–Zehnder interferometer array

Implementation of optical neural network based on Mach–Zehnder interferometer array

MXene‐Based Broadband Ultrafast Nonlinear Activator for Optical Computing

Dual Optical Frequency Comb Neuron: Co‐Developing Hardware and Algorithm

Contact Info

Product

Resources

About