All-optical diffraction neural network based on microstructure array in optical fiber

Abstract

The invention discloses an all-optical diffraction neural network based on a microstructure array in an optical fiber, and belongs to the technical field of photoelectric calculation. A neural network-based function is executed on a coherent optical signal through modulation of a multi-layer microstructure array in an optical fiber at a light velocity, an effective, rapid and low-occupancy mode for realizing a machine learning task is created, and a large-scale neural network is efficiently realized by using coherent light in a low-power-consumption and low-space-occupancy mode.

Description

Technical Field

[0001] The present invention belongs to the technical field of photonic computing, and more specifically, relates to an all-optical diffraction neural network based on a microstructured array in an optical fiber. Background Art

[0002] An artificial neural network is a neural network that mimics the human brain and aims to implement machine learning techniques similar to artificial intelligence. It processes data by imitating the collaborative working mode of biological neurons. With the advent of the big data era, artificial neural networks have been widely applied in fields such as medical image analysis and semantic recognition. However, since electrical signals are vulnerable to interference, the processing speed is proportional to the energy consumption and the floor area, and traditional neural networks based on circuits are limited by devices such as CPUs and GPUs in the circuit, there is still great room for improvement in the processing speed and processing energy consumption of traditional electrical neural networks when dealing with extremely high data volumes required in data centers and the like. Currently, all-optical neural networks have been developed, which can use optical signals as carriers and implement all-optical machine learning by designing the optical paths of passive devices such as appropriate spatial light modulators and phase plates. Although this modulation method avoids the problems of high power consumption and slow processing speed of electrical neural networks, and can execute some functions of the neural network at the speed of light after the system is arranged, the construction of the optical path requires multiple devices, and has high requirements for optical path design, arrangement space, and calibration, which is not conducive to the integration and industrialization of optical systems. Summary of the Invention

[0003] Aiming at the defects of the prior art, the purpose of the present invention is to provide an all-optical diffraction neural network based on a microstructured array in an optical fiber, aiming to solve the problems of too low miniaturization degree and complex device composition of all-optical neural networks.

[0004] To achieve the above purpose, the present invention provides an all-optical diffraction neural network based on a microstructured array in an optical fiber, including an optical input module, a neural network module, and an information acquisition module. The neural network module includes a computable optical fiber, and the core of the computable optical fiber includes multiple layers of microstructured arrays. Each layer of microstructured array is used to independently perform phase modulation on the coherent optical signal propagating in the core. The optical input module is used to provide a coherent optical signal carrying input information. The coherent optical signal is phase-modulated by the neural network module, and the coherent optical signal after phase modulation is collected by the information acquisition module, and the characteristics of the output signal after phase modulation are analyzed and the information carried by the coherent light is decoded.

[0005] Preferably, the optical input module includes a light source and a spatial light modulator. The light source is used to generate coherent light, and the coherent light is modulated by the spatial light modulator to carry input information, thereby obtaining a coherent optical signal carrying input information.

[0006] Preferably, the all-optical diffraction neural network further includes a microstructure processing system, which is used to write, modify, and eliminate the multi-layer microstructure array in the computable optical fiber.

[0007] Preferably, the microstructure array is a three-dimensional array composed of multi-layer circular planar arrays of refractive index waveguides uniformly distributed along the light propagation direction in the optical fiber.

[0008] Preferably, the information acquisition module is a photodetector. Preferably, the microstructure processing module includes a femtosecond laser and a three-dimensional scanning galvanometer dynamic focusing system. The femtosecond laser is responsible for providing a 1030 nm wavelength laser with a repetition frequency of 100 - 500 kHz, a maximum power of 10 W, and a pulse width of 300 fs - 1 ps. The three-dimensional scanning galvanometer dynamic focusing system is responsible for deflecting and focusing the laser at the processing site.

[0009] Furthermore, in an embodiment of the present invention, the microstructure processing system is implemented by a femtosecond laser and a three-dimensional scanning galvanometer dynamic focusing system, and the processing parameters of the microstructure array in the optical fiber core are optimized by a deep learning method.

[0010] Furthermore, the computable optical fiber is an OM1 standard multimode optical fiber, supporting multi-band communication at 850 nm and 1300 nm.

[0011] Furthermore, in an embodiment of the present invention, the distance and quantity between the microstructure arrays in the multimode optical fiber core are obtained through pre-training, including: establishing a simulation model, inputting a training set and a test set according to task requirements, optimizing the structure during the training process through the error backpropagation algorithm, and adjusting the processing parameters of the microstructure array and the physical distance between each module; after the simulation design is completed, use a femtosecond laser and a three-dimensional scanning galvanometer dynamic focusing system for physical manufacturing to execute specific neural network functions.

[0012] Through the above technical solution conceived by the present invention, compared with the prior art, the following

[0013] beneficial effects can be achieved:

[0014] 1. The all-optical diffraction neural network based on the microstructural array in optical fiber according to the embodiments of the present invention performs functions based on neural networks on coherent optical signals through modulation by a multi-layer microstructural array in the optical fiber at the speed of light, creating an effective, fast, and low-occupation way to implement machine learning tasks, and efficiently implementing large-scale neural networks with coherent light in a low-power and low-space-occupation manner. Using this neural network can enrich the functions of existing optical systems while reducing the demand for electrical devices, reducing the response time and increasing the processing speed. For example, a section of computable optical fiber is added to the input surface of an existing photodetector to enable eigenvalue extraction of the input optical field in advance, simplifying subsequent analysis steps.

[0015] 2. The present invention uses a microstructural array to reduce or even completely replace the spatial light modulator required in the optical path except for the optical fiber, while normally implementing the functions of the neural network, thereby achieving the reduction of the space required for arranging the all-optical neural network and the reduction of the calibration requirements. The present invention can improve the integration degree of the all-optical neural network model, and can realize the light weight and high responsiveness of devices such as photodetectors and modulators that require the participation of electrical devices in control, making the application of a pure optical system in the fields of image recognition and optical sensing possible. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 is a schematic structural diagram of the all-optical diffraction neural network based on the microstructural array in optical fiber according to an embodiment of the present invention.

[0017] Figure 2 is a schematic diagram of the distribution of the microstructure in the core.

[0018] Figure 3 is a schematic diagram of an example for manufacturing a computable optical fiber.

[0019] Figure 4 is a flowchart of the method for implementing the machine learning function by the computable optical fiber.

[0020] Figure 5 is an application of the all-optical diffraction neural network based on the microstructural array in optical fiber according to an embodiment of the present invention in the design of a photodetector.

[0021] Figure 6 is a schematic diagram of the processing effect of the computable optical fiber on the input signal light. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0022] In order to make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0023] The all-optical diffraction neural network based on the microstructured array in optical fiber proposed according to an embodiment of the present invention will be described below with reference to the accompanying drawings. First, the all-optical diffraction neural network based on the microstructured array in optical fiber proposed according to an embodiment of the present invention will be described with reference to the accompanying drawings.

[0024] Figure 1 It is a schematic structural diagram of the all-optical diffraction neural network based on the microstructured array in optical fiber according to an embodiment of the present invention.

[0025] As Figure 1 shown, the all-optical diffraction neural network 100 based on the microstructured array in optical fiber includes: an optical input module 110, a neural network module 120, and an information acquisition module 130.

[0026] Among them, the optical input module 110 is used to provide a modulated optical signal. The neural network module 120 is used to transform, extract, and compress the coherent signal. The information acquisition module 130 is used to receive the output signal of the neural network module and generate a processing result according to the output signal.

[0027] The all-optical diffraction neural network 100 based on the microstructured array in optical fiber will be elaborated in detail below.

[0028] Further, in an embodiment of the present invention, as Figure 1 shown, the optical input module 110 includes: a light source 111, a spatial light modulator 112.

[0029] Among them, the light source 111 is a single-longitudinal-mode solid-state laser with a working wavelength of 532 nm, which is used to provide uniform coherent light. The spatial light modulator 112 is used to modulate the coherent light and convert it into coherent light carrying input information.

[0030] Further, in an embodiment of the present invention, the optical input module 110 is further used to convert the information to be processed into the spatial distribution of the optical field by using the spatial light modulator 112.

[0031] Further, in an embodiment of the present invention, the neural network module 120 includes: a computable optical fiber 121, a microstructured processing system 122.

[0032] Among them, the computable optical fiber 121 is an OM1 multimode optical fiber with a core radius of 62.5 / 125 um before processing. The OM1 multimode optical fiber can be regarded as a scattering medium under long-distance transmission. The present invention extracts the characteristic values of the scattered light spot by using multi-layer phase modulation. The core of the computable optical fiber 121 contains a multi-layer microstructure array, and each layer of the microstructure array independently performs phase modulation on the coherent optical signal propagating in the core. The microstructure processing system 122 is used to write, modify, and eliminate the microstructures in the core of the computable optical fiber 121 by using a 1030 nm femtosecond laser. The microstructure array is composed of refractive index waveguide type structural units. At a specific position, the microstructure shape processed by specific parameters is an optical waveguide. The material of the optical waveguide is the same as the chemical composition of the unprocessed core, but due to the change in structure, the molecular density in the processing area increases or decreases, and the refractive index of the processing area changes compared with that of the unprocessed area, forming a waveguide. An optical path difference is generated between the light passing through the processing area in the core and the light passing through the unprocessed area. When propagating to the next modulation plane, the light with different optical paths interferes with each other. After the light field distribution of the input plane is modulated by multiple layers of regions, the light passing through each pixel point interferes with each other multiple times to achieve the target light field distribution at the output plane. The specific processing parameters are optimized by using the random gradient descent algorithm with the difference between the normalized autocorrelation coefficient of the binary image of the output light spot and the target light spot of the computable optical fiber 121 and one as the loss function during training. The specific distribution of the waveguide-type microstructures in the core can refer to Figure 2 .

[0033] Further, in an embodiment of the present invention, the information acquisition module 130 includes: a photodetector 131.

[0034] Among them, the photodetector 131 is used to receive the output signal of the neural network module and convert the optical signal into an electrical signal.

[0035] Specifically, the function of the photodetector 131 is to use the photodetector to convert the optical field information of each pixel point on the receiving surface of the photodetector into an electrical signal for data processing and display.

[0036] The all-optical diffraction neural network 100 of the embodiment of the present invention can improve the integration degree of the optical neural network, expand the application scenarios of the optical neural network, and enable the optical neural network to better complete the machine learning tasks in practice, especially to be able to complete natural scene image recognition, big data analysis processing, and computing tasks.

[0037] The all-optical diffraction neural network 100 of the embodiment of the present invention aims to reduce the use of additional passive optical components such as spatial light modulators and wave plates, reduce the operation process of calibrating between different devices, thereby reducing the overall loss rate of the optical path, and thus reducing the requirements for calibration of the space required for arranging the all-optical neural network, especially reducing the process accuracy requirements for alignment during assembly.

[0038] The following will further elaborate on the working process of the all-optical diffraction neural network 100 based on the microstructured array in the optical fiber in conjunction with Figure 3 the following.

[0039] As Figure 3 shown, the coherent optical signal is irradiated onto the spatial light modulator 112 through the light source 111, and the coherent light is converted into coherent light carrying information through the action of the modulator. The coherent light diffracts when entering the neural network module 120, enters the core from the input end face of the computable optical fiber 121, passes through the multi-layer microstructured array written by the microstructured processing system 122, and performs processing work such as feature extraction and information compression through the modulation of the spatial phase distribution. The output optical signal is received by the photodetector 132.

[0040] The following will introduce in detail the establishment process of the all-optical diffraction neural network 100 based on the microstructured array in the optical fiber of the present invention through specific embodiments. The parameters of the all-optical diffraction neural network 100 in the embodiments of the present invention are optimized by establishing a simulation model and using deep learning methods. Specifically, Figure 4 The following is a flowchart of a method for implementing a machine learning function according to an embodiment of the present invention. The specific process is as follows:

[0041] S1. Establish a numerical simulation model of the optical system of the all-optical diffraction neural network based on the microstructured array in the optical fiber.

[0042] In the optical input module, the all-optical diffraction neural network in the embodiments of the present invention uses the coherent light emitted from the light source as the signal light and inputs it to the receiving surface of the spatial light modulator. After passing through the spatial light modulator, the output coherent signal light has the following relationship E out (x,y) = f(I in (x,y), x, y), where I in (x,y) is the intensity of the coherent light input on the spatial light modulator, and E out (x,y) is the coherent light field output at this point. f(I, x, y) is the corresponding relationship between the output light field at a selected position on the spatial light modulator and the input light intensity at this point, which is determined according to the properties of the photorefractive material itself and the parameters of the recording light and the reading light. The output coherent signal light E out (x,y) becomes the input of the neural network module

[0043] In the neural network module, the input-output relationship of each section of the optical fiber core and each layer of the microstructured array can be expressed as a complex transmittance matrix t(x,y). The output light field passing through the core or the microstructured array is the product of the input light field and the transmittance function. The transmittance function is calculated using the Fresnel-Kirchhoff diffraction theorem. The output of the last layer of the microstructured array is collected by the information acquisition module after propagation.

[0044] In the information acquisition module, the detector receives the intensity or phase information of the above output optical field for storage or further processing.

[0045] S2. Use the deep learning method to simulate the optimal structure of the all-optical diffraction neural network based on the microstructured array in the optical fiber.

[0046] Based on the above simulation model, a deep learning network is established, with the image to be processed as the input and the correct result of the target task as the ground truth, and a suitable training set, validation set, and test set are constructed. In this step, the complex transmittance matrix is the optimization objective, and the optimizable variables are the size of a single microstructure and the relative distribution of multiple microstructures. The weighted average of the difference between the target complex transmittance matrix and the actual complex transmittance matrix is used as the loss function. The random gradient descent algorithm is used to calculate the gradient of the optimizable variables with respect to the input, and the next round is updated according to the gradient. By adjusting hyperparameters such as the distance between arrays and the number of arrays, the best simulation results are obtained.

[0047] S3. Perform preliminary processing on the optical fiber to be processed according to the simulated structure obtained by simulation optimization.

[0048] Use a femtosecond laser and a three-dimensional scanning galvanometer dynamic focusing system to physically write a microstructured array in the optical fiber core. Write the microstructured array at each position of the optical fiber core according to the simulation model, and the function of the all-optical diffraction neural network based on the microstructured array in the optical fiber can be initially realized.

[0049] S4. Optimize the processing parameters of the preliminarily processed computable optical fiber.

[0050] Use a spatial light modulator to modulate the optical field distribution of the input light, use a photodetector to obtain the optical field distribution of the output light, and calculate the actual complex transmittance matrix. In this step, the complex transmittance matrix is the optimization objective, and the optimizable variable is the matrix describing the processing parameters. The weighted average of the difference between the target complex transmittance matrix and the actual complex transmittance matrix is used as the loss function. After each processing, the random gradient descent algorithm is used to calculate the gradient of the optimizable variables with respect to the input, and the next round of processing parameters is updated according to the gradient. By adjusting hyperparameters such as the processing light source power and the debugging light source wavelength, the best optimization results are obtained.

[0051] The following describes the application scenario of the all-optical diffraction neural network based on the microstructured array in the optical fiber according to an embodiment of the present invention with reference to the accompanying drawings.

[0052] Figure 5 It is an application of the all-optical diffraction neural network based on the microstructured array in the optical fiber in the design of a photodetector according to an embodiment of the present invention.

[0053] As Figure 5As shown, the application of the all-optical diffraction neural network based on the microstructure array in the optical fiber to the photodetector system, the photodetector system includes: an input module, a computable optical fiber, a photodetector and a data processing module.

[0054] Among them, the input module can provide a coherent optical signal to be measured carrying information, providing a light source and a signal for the system. The computable optical fiber is used to preprocess the coherent optical signal to be measured or extract eigenvalue, such as Figure 6 As shown, it can convert the complex input optical field distribution into an easily analyzable and processable focused spot according to the learning result of the neural network. The photodetector is used to receive the optical field distribution after the signal processed by the computable optical fiber and convert it into an electrical signal. The data processing module is used for display or further analysis. The application of the embodiment of the present invention can filter and discard the useless information in the signal, make the subsequent calculation examples reasonably distributed to the effective information, reduce the amount of data processed by electrical equipment, and improve the response speed of the photodetector system.

[0055] Furthermore, in the application of the embodiment of the present invention, the computable optical fiber filters the useless information in the optical field distribution of the signal to be measured, and uses information compression and encoding methods to convert the actual optical field into an easily processable optical field, reducing the requirement for the accuracy of the photodetector and simplifying the requirement for computing power of the whole system.

[0056] Specifically, the computable optical fiber filters the useless information in the optical field distribution of the signal to be measured, and classifies the complex actual optical field distribution into smaller focused spots according to the learning result of the neural network. The requirement for the used photodetector can be reduced to complete the recognition of the position of the focused spot, simplifying the requirement for computing power of the whole system and reducing the cost of the photodetector system.

[0057] In summary, the all-optical diffraction neural network of the embodiment of the present invention is different from the traditional optical neural network. The all-optical diffraction neural network of the embodiment of the present invention reduces the number of traditional passive optical components by jointly adopting the processing of the microstructure array in the optical fiber core, so that all the calibration and layout design tasks of the all-optical diffraction neural network can be run in a more convenient way, while reducing the equipment requirements of the high-precision measurement system for fixed purposes, greatly improving the integration degree of the system, expanding the application range, and being widely applicable to multiple fields such as computing, sensing, modulation, and communication.

[0058] It is easy for those skilled in the art to understand that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. An all-optical diffraction neural network based on a microstructure array in an optical fiber, characterized in that: include: An optical input module, a neural network module and an information acquisition module, wherein the neural network module includes a computable optical fiber, and the core of the computable optical fiber includes a multi-layer microstructure array, and each layer of the microstructure array is used to independently phase-modulate the coherent light signal propagating in the core; the optical input module is used to provide a coherent light signal carrying input information, and the coherent light signal is phase-modulated by the neural network module. The coherent light signal after phase modulation is collected by the information acquisition module, and the characteristics of the output signal after phase modulation are analyzed and decoded to obtain the information carried by the coherent light.

2. The all-optical diffractive neural network according to claim 1, characterized in that: The optical input module includes a light source and a spatial light modulator. The light source is used to generate coherent light. The coherent light is modulated by the spatial light modulator to carry input information, thereby obtaining a coherent light signal carrying the input information.

3. The all-optical diffractive neural network according to claim 1, characterized in that: Also included is a microstructure processing system for writing, modifying and erasing multi-layer microstructure arrays in a computable optical fiber.

4. The all-optical diffractive neural network according to claim 1, characterized in that: The microstructure array is a three-dimensional array composed of multiple layers of uniformly distributed refractive index waveguide circular plane arrays along the propagation direction of light.

5. The all-optical diffractive neural network according to claim 3, characterized in that: The microstructure processing module includes a femtosecond laser and a three-dimensional scanning galvanometer dynamic focusing system. The femtosecond laser is used to provide laser light, and the three-dimensional scanning galvanometer dynamic focusing system is used to deflect and focus the laser light at a processing site.

6. The all-optical diffractive neural network according to claim 1, characterized in that: The information acquisition module is a photoelectric detector.

7. The all-optical diffractive neural network according to claim 1, characterized in that: The computable optical fiber is an OM1 standard multimode optical fiber.

Images