Optical Neural Network Research Articles

End-to-End Optimization for a Compact Optical Neural Network Based on Nanostructured 2 × 2 Optical Processors

End-to-End Optimization for a Compact Optical Neural Network Based on Nanostructured 2 × 2 Optical Processors

Survey on Activation Functions for Optical Neural Networks

Integrated photonics arises as a fast and energy-efficient technology for the implementation of artificial neural networks (ANNs). Indeed, with the growing interest in ANNs, photonics shows great promise to overcome current limitations of electronic-based implementation. For example, it has been shown that neural networks integrating optical matrix multiplications can potentially run two orders of magnitude faster than their electronic counterparts. However, the transposition in the optical domain of the activation functions, which is a key feature of ANNs, remains a challenge. There is no direct optical implementation of state-of-the-art activation functions. Currently, most designs require time-consuming and power-hungry electro-optical conversions. In this survey, we review both all-optical and opto-electronic activation functions proposed in the state-of-the-art. We present activation functions with their key characteristics, and we summarize challenges for their use in the context of all-optical neural networks. We then highlight research directions for the implementation of fully optical neural networks.

Simulation of information transmission of IoT communication nodes based on fuzzy hierarchical optical neural network

Simulation of information transmission of IoT communication nodes based on fuzzy hierarchical optical neural network

Aided diagnosis of thyroid nodules based on an all-optical diffraction neural network.

Thyroid cancer is the most common malignancy in the endocrine system, with its early manifestation being the presence of thyroid nodules. With the advantages of convenience, noninvasiveness, and a lack of radiation, ultrasound is currently the first-line screening tool for the clinical diagnosis of thyroid nodules. The use of artificial intelligence to assist diagnosis is an emerging technology. This paper proposes the use optical neural networks for potential application in the auxiliary diagnosis of thyroid nodules. Ultrasound images obtained from January 2013 to December 2018 at the Institute and Hospital of Oncology, Tianjin Medical University, were included in a dataset. Patients who consecutively underwent thyroid ultrasound diagnosis and follow-up procedures were included. We developed an all-optical diffraction neural network to assist in the diagnosis of thyroid nodules. The network is composed of 5 diffraction layers and 1 detection plane. The input image is placed 10 mm away from the first diffraction layer. The input of the diffractive neural network is light at a wavelength of 632.8 nm, and the output of this network is determined by the amplitude and light intensity obtained from the detection region. The all-optical neural network was used to assist in the diagnosis of thyroid nodules. In the classification task of benign and malignant thyroid nodules, the accuracy of classification on the test set was 97.79%, with an area under the curve value of 99.8%. In the task of detecting thyroid nodules, we first trained the model to determine whether any nodules were present and achieved an accuracy of 84.92% on the test set. Our study demonstrates the potential of all-optical neural networks in the field of medical image processing. The performance of the models based on optical neural networks is comparable to other widely used network models in the field of image classification.

Evanescent coupling of nonlinear integrated cavities for all-optical reservoir computing

We consider theoretically a network of evanescently coupled optical microcavities to implement a space-multiplexed optical neural network in an integrated nanophotonic circuit. Nonlinear photonic network integrations based on evanescent coupling ensure a highly dense integration, reducing the chip footprint by several orders of magnitude compared to commonly used designs based on long waveguide connections while allowing the processing of optical signals with bandwidth in a practical range. Different nonlinear effects inherent to such microcavities are studied for realizing an all-optical autonomous computing substrate based on the reservoir computing concept, and their contribution to computing performance is demonstrated. We provide an in-depth analysis of the impact of basic microcavity parameters on the computational metrics of the system, namely, the dimensionality and the consistency. Importantly, we find that differences between frequencies and bandwidths of supermodes formed by the evanescent coupling are the determining factor of the reservoir’s dimensionality and scalability. The network’s dimensionality can be improved with frequency-shifting nonlinear effects such as the Kerr effect, while two-photon absorption has the opposite effect. Finally, we demonstrate in simulation that the proposed reservoir is capable of solving the Mackey–Glass prediction and the optical signal recovery tasks at gigahertz timescale.

Harnessing exciton-polaritons for digital computing, neuromorphic computing, and optimization [Invited

Polaritons are quasiparticles resulting from the strong quantum coupling of light and matter. Peculiar properties of polaritons are a mixture of physics usually restricted to one of these realms, making them interesting for study not only from the fundamental point of view but also for applications. In recent years, many studies have been devoted to the potential use of exciton-polaritons for computing. Very recently, it has been shown experimentally that they can be harnessed not only for digital computing but also for optical neural networks and for optimization related to hard computational problems. Here, we provide a brief review of recent studies and the most important results in this area. We focus our attention, in particular, on the emerging concepts of non-von-Neumann computing schemes and their realizations in exciton-polariton systems.

Microcomb-Driven Optical Convolution for Car Plate Recognition

The great success of artificial intelligence (AI) calls for higher-performance computing accelerators, and optical neural networks (ONNs) with the advantages of high speed and low power consumption have become competitive candidates. However, most of the reported ONN architectures have demonstrated simple MNIST handwritten digit classification tasks due to relatively low precision. A microring resonator (MRR) weight bank can achieve a high-precision weight matrix and can increase computing density with the assistance of wavelength division multiplexing (WDM) technology offered by dissipative Kerr soliton (DKS) microcomb sources. Here, we implement a car plate recognition task based on an optical convolutional neural network (CNN). An integrated DKS microcomb was used to drive an MRR weight-bank-based photonic processor, and the computing precision of one optical convolution operation could reach 7 bits. The first convolutional layer was realized in the optical domain, and the remaining layers were performed in the electrical domain. Totally, the optoelectronic computing system (OCS) could achieve a comparable performance with a 64-bit digital computer for character classification. The error distribution obtained from the experiment was used to emulate the optical convolution operation of other layers. The probabilities of the softmax layer were slightly degraded, and the robustness of the CNN was reduced, but the recognition results were still acceptable. This work explores an MRR weight-bank-based OCS driven by a soliton microcomb to realize a real-life neural network task for the first time and provides a promising computational acceleration scheme for complex AI tasks.

Data-Driven Modeling of Mach-Zehnder Interferometer-Based Optical Matrix Multipliers

Photonic integrated circuits are facilitating the development of optical neural networks, which have the potential to be both faster and more energy efficient than their electronic counterparts since optical signals are especially well-suited for implementing matrix multiplications. However, accurate programming of photonic chips for optical matrix multiplication remains a difficult challenge. Here, we describe both simple analytical models and data-driven models for offline training of optical matrix multipliers. We train and evaluate the models using experimental data obtained from a fabricated chip featuring a Mach-Zehnder interferometer mesh implementing 3-by-3 matrix multiplication. The neural network-based models outperform the simple physics-based models in terms of prediction error. Furthermore, the neural network models are also able to predict the spectral variations in the matrix weights for up to 100 frequency channels covering the C-band. The use of neural network models for programming the chip for optical matrix multiplication yields increased performance on multiple machine learning tasks.

On chip optical neural networks based on MMI microring resonators for image classification

We propose a new on-chip optical neural network (OONN) based on multimode interference-microring resonators (MMI-RRs). The suggested structure eliminates the need for wavelength division multiplexers (WDM) to create an optical neuron on a single chip. New microring resonator structure based on 4×4 MMI coupler with a size of 24µm × 2900 µm is used for the basic elements of the computation matrix, as a result a higher bandwidth and free spectral range (FSR) can be achieved. The Si3N4 platform along with the graphene sheet is designed to modulate the signals and weights of the neural networks at a very high speed. The Si3N4 can provide wide range of operating wavelengths and can work directly with the wavelengths of color images. The structure's benefits include rapid computing speed, little loss, and the ability to handle both positive and negative values. The OONN has been applied to the MNIST dataset with a speed faster than 2.8 to 14x times compared with the conventional GPU methods.

Deep learning for highly efficient curvature recognition using fiber scattering speckles

A flexible fiber-optic sensor enabled by deep learning is proposed and experimentally demonstrated for highly efficient curvature sensing application. This sensing modulation system combines a deep optical neural network based on a small training dataset, aiming to simplify speckle data capture and sensor model evaluation. The multimode fiber concatenated with a section of single stress-applying fiber serves as a sensing unit as well as an image transport medium. A type of hybrid scattering speckle images is collected and employed to provide more freedom to identify the bending curvature with and without external disturbances. In a perturbed environment, the trained optical classification model is suitable for the speckle dataset recognition with high accuracy rate of 98.3%. Moreover, the deep-learning-enabled fiber curvature sensor shows great potential for practical applications in real-time structural safety test, including studies on health monitoring of infrastructure equipment and aerospace wings.

Time-stretch optical neural network with time-division multiplexing

As factors such as power consumption and bandwidth constrain the development of electronic computing in deep learning, optical neuromorphic computing has the potential to meet the growing computational demands of artificial intelligence, presenting various opportunities and challenges. In this paper, we propose an all-optical neural network through a combination of time-stretching and time-division multiplexing that can be used for the classification of multiple datasets. The system modulates a large number of parameters into chirp optical pulses by time-stretching and increases operational efficiency in conjunction with time-division multiplexing. This all-optical neural network system proposed is capable of operating at the speed of 1 Tera-OPS/s and can achieve data transfer rate in excess of 800 Gbit/s. The system is tested by the feedforward neural network, and the computational simulation realized very encouraging and considerable results.

Smart explainable artificial intelligence for sustainable secure healthcare application based on quantum optical neural network

Smart explainable artificial intelligence for sustainable secure healthcare application based on quantum optical neural network

Low-threshold all-optical nonlinear activation function based on injection locking in distributed feedback laser diodes.

We experimentally demonstrate an all-optical nonlinear activation unit based on the injection-locking effect of distributed feedback laser diodes (DFB-LDs). The nonlinear carrier dynamics in the unit generates a low-threshold nonlinear activation function with optimized operating conditions. The unit can operate at a low threshold of -15.86 dBm and a high speed of 1 GHz, making it competitive among existing optical nonlinear activation approaches. We apply the unit to a neural network task of solving the second-order ordinary differential equation. The fitting error is as low as 0.0034, verifying the feasibility of our optical nonlinear activation approach. Given that the large-scale fan-out of optical neural networks (ONNs) will significantly reduce the optical power in one channel, our low-threshold scheme is suitable for the development of high-throughput ONNs.

Implementation of energy-efficient convolutional neural networks based on kernel-pruned silicon photonics.

Silicon-based optical neural networks offer the prospect of high-performance computing on integrated photonic circuits. However, the scalability of on-chip optical depth networks is restricted by the limited energy and space resources. Here, we present a silicon-based photonic convolutional neural network (PCNN) combined with the kernel pruning, in which the optical convolutional computing core of PCNN is a tunable micro-ring weight bank. Our numerical simulation demonstrates the effect of weight mapping accuracy on PCNN performance and we find that the performance of PCNN decreases significantly when the weight mapping accuracy is less than 4.3 bits. Additionally, the experimental demonstration shows that the accuracy of the PCNN on the MNIST dataset has a slight loss compared to the original CNN when 93.75 % of the convolutional kernels are pruned. By making use of kernel pruning, the energy saved by a convolutional kernel removal is about 202.3 mW, and the overall energy saved has a linear relationship with the number of kernels removed. The methodology is scalable and provides a feasible solution for implementing faster and more energy-efficient large-scale optical convolutional neural networks on photonic integrated circuits.

Electrically programmable phase-change photonic memory for optical neural networks with nanoseconds in situ training capability

Electrically programmable phase-change photonic memory for optical neural networks with nanoseconds in situ training capability

Nonlinear absorption of 2D materials and their application in optical neural networks

In recent years, optical neural network (ONN) research has blossomed due to the outstanding advantage of energy consumption and computing property. Regrettably, nonlinear processing in the optical domain remains a huge challenge. The optical characteristics of 2D material, particularly related to saturable absorption (SA), have enabled nonlinear operation. Here, we discuss the SA models with various categories and their application in ONNs. A feedforward artificial neural network was built for handwritten digit recognition to illustrate the feasibility of SA features as nonlinear mapping. For comparison, ONNs without the assistance of the activation function were used as a benchmark to examine the capability of the nonlinear models. A simulation shows that the accuracy of digit classification ranged from 86% to 95%, depending on the nonlinearity of the mediums. This work offers an optical nonlinear unit selection guideline to explore ONNS.

Single-shot optical neural network.

Analog optical and electronic hardware has emerged as a promising alternative to digital electronics to improve the efficiency of deep neural networks (DNNs). However, previous work has been limited in scalability (input vector length K ≈ 100 elements) or has required nonstandard DNN models and retraining, hindering widespread adoption. Here, we present an analog, CMOS-compatible DNN processor that uses free-space optics to reconfigurably distribute an input vector and optoelectronics for static, updatable weighting and the nonlinearity-with K ≈ 1000 and beyond. We demonstrate single-shot-per-layer classification of the MNIST, Fashion-MNIST, and QuickDraw datasets with standard fully connected DNNs, achieving respective accuracies of 95.6, 83.3, and 79.0% without preprocessing or retraining. We also experimentally determine the fundamental upper bound on throughput (∼0.9 exaMAC/s), set by the maximum optical bandwidth before substantial increase in error. Our combination of wide spectral and spatial bandwidths enables highly efficient computing for next-generation DNNs.

Optical Diffractive Convolutional Neural Networks Implemented in an All-Optical Way

Optical neural networks can effectively address hardware constraints and parallel computing efficiency issues inherent in electronic neural networks. However, the inability to implement convolutional neural networks at the all-optical level remains a hurdle. In this work, we propose an optical diffractive convolutional neural network (ODCNN) that is capable of performing image processing tasks in computer vision at the speed of light. We explore the application of the 4f system and the diffractive deep neural network (D2NN) in neural networks. ODCNN is then simulated by combining the 4f system as an optical convolutional layer and the diffractive networks. We also examine the potential impact of nonlinear optical materials on this network. Numerical simulation results show that the addition of convolutional layers and nonlinear functions improves the classification accuracy of the network. We believe that the proposed ODCNN model can be the basic architecture for building optical convolutional networks.

C-DONN: compact diffractive optical neural network with deep learning regression.

A new method to improve the integration level of an on-chip diffractive optical neural network (DONN) is proposed based on a standard silicon-on-insulator (SOI) platform. The metaline, which represents a hidden layer in the integrated on-chip DONN, is composed of subwavelength silica slots, providing a large computation capacity. However, the physical propagation process of light in the subwavelength metalinses generally requires an approximate characterization using slot groups and extra length between adjacent layers, which limits further improvements of the integration of on-chip DONN. In this work, a deep mapping regression model (DMRM) is proposed to characterize the process of light propagation in the metalines. This method improves the integration level of on-chip DONN to over 60,000 and elimnates the need for approximate conditions. Based on this theory, a compact-DONN (C-DONN) is exploited and benchmarked on the Iris plants dataset to verify the performance, yielding a testing accuracy of 93.3%. This method provides a potential solution for future large-scale on-chip integration.

Bioinspired Artificial Visual System Based on 2D WSe2 Synapse Array

AbstractMachine vision systems that capture images for visual inspection and recognition tasks must be able to perceive, memorize, and compute any color scene. To achieve this, most of the current visual systems use circuits and algorithms which may reduce efficiency and increase complexity. Herein, a 2D semiconductor tungsten diselenide (WSe2)‐based phototransistor that successfully demonstrates an artificial vision system integrating the processing capability of visual information sensing memory, is reported. Furthermore, based on a 6 × 6 fabricated retinal perception array, artificial visual information sensing memory and processing system are proposed to perform image recognition tasks, which can avoid the time delay and energy consumption caused by data conversion and movement. On the other hand, highly linear symmetric synaptic plasticity can be achieved based on the modulation of carrier types in WSe2 transistors with different thicknesses, facilitating the high level of training and inference accuracy for artificial neural networks. Last, through training and inference simulations, the feasibility of the hybrid synapses for optical neural networks (ONN) is demonstrated.