All-optical neural network system and optical information processing method

Abstract

This application provides an all-optical neural network system and an optical information processing method. The system includes: a pulsed laser source for emitting a pulsed laser beam; a digital micromirror device for loading input image information and encoding the input image information onto the laser beam irradiating its surface; a phase modulation device for phase modulation of the laser beam carrying the input image information emitted from the surface of the digital micromirror device to form an optical signal with a predetermined intensity distribution on the output plane; a fiber delay line array device, including a fiber delay line array and a combiner, for applying different time delays to the optical signals with the predetermined intensity distribution and combining them for output; a single-pixel high-speed detector for receiving the optical signal output from the fiber delay line array device and converting the optical signal into an electrical signal at high speed; and a signal acquisition and processing module for acquiring the electrical signal to obtain one-dimensional pulse time-domain waveform signal data, and then obtaining the input image information.

Description

Technical Field

[0001] This application relates to the field of intelligent optical computing technology, and in particular to an all-optical neural network system and an optical information processing method. Background Technology

[0002] With the rapid development of technologies such as artificial intelligence and big data, the demand for high-performance, low-latency intelligent computing systems is increasing daily. However, current mainstream computing architectures (such as GPUs (Graphics Processing Units) and TPUs (Tensor Processing Units)) are based on the traditional von Neumann architecture, facing a structural bottleneck of "separation of computation and storage," which cannot effectively meet the demands for ultra-high-speed processing and high energy efficiency. Optical Neural Networks (ONNs) are considered an important direction for breaking through the bottlenecks of traditional electronic computing due to their inherent high speed and parallelism. Optical computing does not require charge transport during execution and can complete matrix multiplication, convolution, and other operations in neural networks with bandwidth at the THz level. However, the practical application of optical computing systems is currently still limited by the result reading stage. The output of most existing optical neural networks relies on two-dimensional photoelectric sensors (such as CMOS (Complementary Metal-Oxide-Semiconductor Transistors) and CCD (Charge Coupled Device) cameras) for imaging acquisition. The frame rates of these devices are usually on the order of kHz, which greatly limits the inference speed of the entire system. Even with commercial high-speed cameras, frame rates struggle to break the MHz level, a significant gap compared to the inherent processing speed of optical computing, causing the bottleneck to shift to the output. Some studies have attempted to alleviate this problem using compressed sensing and optical frequency domain coding, but these methods still fail to overcome the structural limitations of two-dimensional readout, resulting in high system complexity, high energy consumption, and poor stability. Summary of the Invention

[0003] The purpose of this application is to provide an all-optical neural network system and an optical information processing method, which can at least partially solve the above-mentioned technical problems existing in the prior art.

[0004] One aspect of this application provides an all-optical neural network system. The system includes: a pulsed laser source for emitting a pulsed laser beam; a digital micromirror device for loading input image information and encoding the input image information onto the laser beam irradiating its surface; a phase modulation device for phase-modulating the laser beam carrying the input image information emitted from the surface of the digital micromirror device to form an optical signal with a predetermined intensity distribution on an output plane; a fiber delay line array device, including a fiber delay line array and a combiner, wherein the fiber delay line array applies different time delays to the optical signal with the predetermined intensity distribution, and the combiner combines and outputs the optical signals with different time delays; a single-pixel high-speed detector for receiving the optical signal output from the fiber delay line array device and converting the optical signal into an electrical signal at high speed; and a signal acquisition and processing module for acquiring the electrical signal output from the single-pixel high-speed detector to obtain one-dimensional pulse time-domain waveform signal data, thereby obtaining the input image information.

[0005] Furthermore, the all-optical neural network system also includes a reflector used to change the path of the laser beam so that the laser beam is incident on the surface of the digital micromirror device at a predetermined angle.

[0006] Furthermore, the all-optical neural network system also includes a 4f system, which is used to filter out the multiple diffraction orders caused by the laser beam passing through the digital micromirror device, and retain only the required diffraction orders to continue propagating forward.

[0007] Furthermore, the all-optical neural network system also includes a beam splitter, which is used to transmit the light field emitted from the 4f system and reflect the light field returned from the phase modulation device perpendicularly to the input light field and output it to the fiber delay line array device.

[0008] Furthermore, the all-optical neural network system also includes a beam expander and a first aperture stop, wherein the beam expander is used to increase the diameter of the laser beam; and the first aperture stop is used to filter the edge of the laser beam after it has been expanded by the beam expander.

[0009] Furthermore, the all-optical neural network system also includes a Fourier transform lens and a second aperture stop, wherein the Fourier transform lens is used to apply a Fourier transform to the output light field of the phase modulation device; and the second aperture stop is used to filter the light intensity distribution of other diffraction orders generated after the Fourier transform.

[0010] Furthermore, the phase modulation device includes a spatial light modulator, which is used to construct a single-layer diffractive optical neural network structure. The phase distribution loaded on the spatial light modulator is obtained by pre-training on a computer using a deep learning algorithm to optimize the two-dimensional light intensity distribution output by the spatial light modulator into multiple discretized spatial focal points. The arrangement of the multiple spatial focal points on the output plane of the spatial light modulator is the same as the arrangement of the fiber delay lines in the fiber delay line array on the input plane of the fiber delay line array device. Each fiber delay line in the fiber delay line array has a different length.

[0011] Furthermore, the phase modulation device includes a plurality of spatial light modulators, which are arranged at equal intervals according to the incident direction of the laser beam.

[0012] Another aspect of this application provides an optical information processing method. The method includes: encoding input image information onto a pulsed laser beam; incident the laser beam carrying the input image information onto a phase modulation device for phase modulation to output an optical signal with a predetermined intensity distribution; applying different time delays to the optical signal with the predetermined intensity distribution and merging the output signals; coupling the merged output optical signal into a single-pixel high-speed photodetector for high-speed photoelectric conversion; and acquiring and analyzing the converted output electrical signal to obtain one-dimensional pulse time-domain waveform signal data, thereby obtaining the input image information.

[0013] Furthermore, the phase modulation device includes a spatial light modulator, and the method further includes: optimizing the phase distribution of the spatial light modulator in advance by constructing a forward propagation algorithm and an error backpropagation algorithm on a computer, and loading the optimized phase distribution onto the spatial light modulator to optimize the two-dimensional light intensity distribution output by the spatial light modulator into multiple discretized spatial focal points.

[0014] Furthermore, the predetermined light intensity distribution includes multiple discrete spatial focal points, and the step of applying different time delays to the optical signals of the predetermined light intensity distribution and merging the outputs includes: incidenting the optical signals of each spatial focal point onto the input ends of each optical fiber delay line in the optical fiber delay line array device, wherein each optical fiber delay line has a different length, and the arrangement of the multiple spatial focal points on the output plane of the spatial light modulator is the same as the arrangement of the optical fiber delay lines in the optical fiber delay line array on the input plane of the optical fiber delay line array device; and merging the optical signals output from the output ends of each optical fiber delay line into a time-domain multiplexed sequence using a combiner in the optical fiber delay line array device.

[0015] Furthermore, the method further includes: controlling the gradient distribution of the lengths of each fiber delay line to generate a predetermined time delay difference between the optical signals passing through each fiber delay line.

[0016] Furthermore, a tunable delay line is provided at the output end of the beam combiner, and the method further includes: using the tunable delay line to perform closed-loop feedback control on the output of the beam combiner to correct the time delay error in real time.

[0017] Furthermore, the method further includes: extracting characteristic peaks passing through each optical fiber delay line from the one-dimensional pulse time-domain waveform signal data using a matched filter bank to form a peak amplitude sequence; comparing each characteristic peak in the peak amplitude sequence with a set detection threshold to obtain the encoded information of the input image information.

[0018] The all-optical neural network system and optical information processing method of one or more embodiments of this application have at least the following beneficial technical effects:

[0019] (1) Eliminate the limitations of two-dimensional detectors: adopt a time-series reading mechanism to avoid frame rate bottlenecks and support data reading at the GHz frequency level;

[0020] (2) Simple structure and low cost: The system only requires a single-pixel high-speed detector, which significantly reduces hardware complexity and cost;

[0021] (3) Strong system stability: The physical structure of the length difference of the fiber delay line is determined, the response consistency is high, and the anti-disturbance ability is strong;

[0022] (4) High scalability: The number and length difference of fiber delay lines can be freely adjusted according to the number of output channels to achieve different precision and resolution;

[0023] (5) Good compatibility: It is suitable for various optical neural network structures and has a universal high-speed output interface function. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the structure of an all-optical neural network system according to an embodiment of this application.

[0025] Figure 2 This is a schematic diagram illustrating the forward propagation of the light field and the backward propagation of the error during the phase distribution training process of a spatial light modulator, according to one embodiment of this application.

[0026] Figure 3 This is a diagram showing the operational test results of an all-optical neural network system according to an embodiment of this application.

[0027] Figure 4 This is a flowchart of an optical information processing method according to an embodiment of this application.

[0028] Figure 5 This is a schematic diagram of a spatiotemporal projection method for optical fields based on an optical fiber delay line array device according to an embodiment of this application.

[0029] Figure 6 This is a schematic diagram of the spatial light intensity distribution used in the light field spatiotemporal projection method of this application.

[0030] Figure 7 To utilize the light field spatiotemporal projection method of this application, a one-dimensional time-domain waveform signal result is obtained by rapidly switching the input image on the digital micromirror device and measuring it with a single-pixel high-speed detector. Detailed Implementation

[0031] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses consistent with some aspects of this application as detailed in the appended claims.

[0032] The all-optical neural network system and optical information processing method of various embodiments of this application will be described in detail below with reference to the accompanying drawings. Unless otherwise specified, the features of the following embodiments and implementation methods can be combined with each other.

[0033] This application provides an all-optical neural network system that can be used to achieve real-time, high-speed, low-energy-consumption, and highly parallel artificial intelligence task inference. Figure 1 A schematic diagram of the structure of an all-optical neural network system 100 according to an embodiment of this application is shown. Figure 1 As shown, an all-optical neural network system 100 according to one embodiment of this application includes a pulsed laser source 101, a digital micromirror device 102, a phase modulation device, a fiber delay line array device 104, a single-pixel high-speed detector 105, and a signal acquisition and processing module 106.

[0034] The pulsed laser source 101 can be used to emit pulsed laser beams. In one embodiment, the pulsed laser source 101 can be a high-repetition-rate picosecond pulsed laser (wavelength of 1064 nm, pulse width of 10 ps, ​​and repetition rate of 10 MHz), which can provide a high-speed pulsed illumination source for images. Of course, in other embodiments, the pulsed laser source 101 can also use a laser with a shorter pulse width, such as a femtosecond laser, to provide finer temporal resolution.

[0035] The digital micromirror device 102 can be used to load input image information and encode it onto a laser beam illuminating its surface. The input image information, for example, is pixel grayscale information of a two-dimensional image. After the laser beam illuminates the surface of the digital micromirror device 102, it is modulated by the device into a light field carrying the input image information, which is emitted perpendicular to the surface of the device. In one embodiment, an ultra-high-speed digital micromirror device 102 can be used, switching the two-dimensional input image at a frequency of up to 10.752 kHz.

[0036] A phase modulation device can be used to phase modulate a laser beam carrying input image information emitted from the surface of a digital micromirror device 102, so as to form an optical signal with a predetermined intensity distribution on the output plane.

[0037] In one embodiment of this application, the phase modulation device may include a spatial light modulator 103. The spatial light modulator 103 is used to construct a single-layer diffractive optical neural network structure. It has a phase modulation function. When a light field passes through the spatial light modulator 103, light fields with different frequencies and spatial components are modulated and diffracted, realizing a function similar to matrix operations. Finally, the modulated light field can form a predetermined light intensity distribution on the output plane of the spatial light modulator 103, representing the output result of the neural network.

[0038] In one embodiment, the phase distribution loaded on the spatial light modulator 103 is obtained by pre-training a deep learning algorithm on a computer. The deep learning algorithm may include, for example, a forward propagation algorithm and an error backpropagation algorithm. That is, a forward propagation algorithm and an error backpropagation algorithm are pre-built on the computer to fit the input information and the output spatial intensity distribution, obtaining the phase parameters of each pixel on the spatial light modulator 103. The forward propagation and error backpropagation processes of the light field are as follows: Figure 2 As shown.

[0039] Specifically, the forward propagation algorithm is built upon Rayleigh-Sommerfeld diffraction theory. In the... The first layer of diffraction layer The location of each neuron can be represented as... The contribution of each neuron to diffraction is determined by the propagation of the incident light wave, and the diffraction effect of each neuron can be expressed as:

[0040]

[0041]

[0042] in, This represents the distance between the neuron and the wave source. It is the wavelength of the incident light. It is the imaginary unit. The first neuron was... The light field modulated by the diffraction layer can be expressed as:

[0043]

[0044] in, Indicates from the Incident light waves of the layer, This represents the set of all neurons in the previous layer. The complex-valued transmission or reflection coefficients of a neuron are given by the following formula:

[0045]

[0046] in, The amplitude parameter represents the light field. The phase parameter representing the light field, in an optical neural network with only phase modulation, It is a constant, and The properties of the diffraction layer can be adjusted, representing learnable network parameters. During training, these parameters can be iteratively optimized using the backpropagation algorithm.

[0047] Specifically, the error backpropagation algorithm is implemented by designing a loss function, where the loss function is a parameter used during model training to measure the difference between the predicted and true values. In one embodiment, it can be implemented using the Mean Squared Error (MSE) formula:

[0048]

[0049] in, Indicates the number of pixels. Representing the The light intensity value of each pixel. Representing the The model predicts values ​​for each pixel. By minimizing the loss function, the phase distribution can be continuously optimized, thereby improving the accuracy of the light intensity distribution output by the neural network.

[0050] In some embodiments, multiple spatial light modulators 103 can be used, and the multiple spatial light modulators 103 are arranged at equal intervals according to the incident direction of the laser beam to realize multi-layer diffraction control of the neural network and achieve more powerful computing capabilities.

[0051] The optimized phase distribution is loaded onto the spatial light modulator 103, thereby optimizing the two-dimensional light intensity distribution output by the spatial light modulator 103 into multiple discretized spatial focal points.

[0052] Of course, it is understood that the phase modulation device of this application is not limited to the spatial light modulator 103. In other embodiments, the spatial light modulator 103 may also be replaced by a liquid crystal phase plate or a 3D printed phase plate. These phase plates are passive and fixed, which can further reduce the energy consumption of the entire system.

[0053] The fiber delay line array device 104 includes a fiber delay line array 1041 and a combiner 1042, such as Figure 5 As shown, the fiber delay line array 1041 includes multiple fiber delay lines, each with a different length. The fiber delay lines can be, for example, multimode fibers, and each fiber delay line represents a different optical signal delay channel. The number of fiber delay lines is the same as the number of spatial focal points output by the spatial light modulator 103. Furthermore, the arrangement of the multiple spatial focal points on the output plane of the spatial light modulator 103 is the same as the arrangement of the fiber delay lines in the fiber delay line array 1041 on the input plane of the fiber delay line array device 104.

[0054] The fiber delay line array 1041 can apply different time delays to the input optical signals with a predetermined light intensity distribution. The combiner 1042 has multiple inputs and one output, and can realize multi-input to single-output optical signal transmission with low loss, thereby combining optical signals with different time delays for output.

[0055] For example, in conjunction with reference Figure 5 In one embodiment of this application, the fiber delay line array 1041 is made of three multimode fibers with different lengths (e.g., 2 m, 4 m, 6 m) (core / cladding size of 200 / 220 μm, NA (Numerical Aperture) of 0.22), and the arrangement of the multimode fibers is similar to... Figure 2 The spatial focal points on the output plane are arranged in the same way. The combiner 1042 has three input terminals and one output terminal, and the average optical signal coupling efficiency from input to output is over 90%. The output terminal of the combiner 1042 can be connected to a 0.20m long multimode optical fiber (core / cladding size of 200 / 220μm, NA of 0.46).

[0056] Of course, in other embodiments, the fiber delay line array device 104 can also be replaced with any optical waveguide device, as long as it can generate different time delays for the optical signal in different delay channels.

[0057] The single-pixel high-speed detector 105 can be used to receive optical signals output from the fiber optic delay line array device 104 and convert the optical signals into electrical signals at high speed. In one embodiment, a photodiode with a 3dB bandwidth of 5GHz can be used as the detector, with a detection wavelength range of 800-1700 nm.

[0058] In one embodiment, the signal acquisition and processing module 106 may, for example, use an oscilloscope with a bandwidth of 32 GHz and a sampling rate of 80 GHz to acquire the electrical signal output from the single-pixel high-speed detector 105, thereby obtaining one-dimensional pulse time-domain waveform signal data. Through further simple processing, the input image information can be inferred. An embodiment of this application provides the inference result for a single frame image.

[0059] In some embodiments, the all-optical neural network system 100 of this application may further include a reflector 111. The reflector 111 may be used to change the path of the laser beam so that the laser beam is incident on the surface of the digital micromirror device 102 at a predetermined angle. In one embodiment, the angle of the reflector 111 is adjusted so that the laser beam is incident on the surface of the digital micromirror device 102 at an incident angle of 12°.

[0060] In some embodiments, the all-optical neural network system 100 of this application may further include a 4f system 112. The 4f system 112 is used to filter out multiple diffraction orders caused by the laser beam passing through the digital micromirror device 102, retaining only the required diffraction orders to continue propagating forward. In one embodiment, the 4f system 112 may be composed of two plano-convex lenses with the same focal length of 50 cm, which, after filtering, retain the +1st order diffraction information of the image.

[0061] In some embodiments, the all-optical neural network system 100 of this application may further include a beam splitter 113. The beam splitter 113 can be used to transmit the light field emitted from the 4f system 112 and reflect the light field subsequently returned from the phase modulation device perpendicularly to the input light field to the fiber delay line array device 104.

[0062] In some embodiments, the all-optical neural network system 100 of this application may further include a beam expander 114. The beam expander 114 can be used to increase the diameter of the laser beam. In one embodiment, the beam expander 114 may be composed of a plano-convex lens with a focal length of 25 cm and a focal length of 100 cm, thereby increasing the diameter of the input laser beam by a factor of 4.

[0063] In some embodiments, the all-optical neural network system 100 of this application may also include a first aperture stop 121. The first aperture stop 121 is used to filter the edge of the laser beam after it has been expanded by the beam expander 114, generating a wavefront that approximates a plane wave, and can retain the laser beam power to the maximum extent.

[0064] In some embodiments, the all-optical neural network system 100 of this application may further include a Fourier transform lens 115. The Fourier transform lens 115 can be used to apply a Fourier transform to the output optical field of the phase modulation device. In one embodiment of this application, the Fourier transform lens 115 may be a plano-convex lens with a focal length of 50 cm, and its distance from both the spatial light modulator 103 and the fiber delay line array device 104 is 50 cm.

[0065] In some embodiments, the all-optical neural network system 100 of this application may further include a second aperture stop 122. The second aperture stop 122 can be used to filter the intensity distribution of other diffraction orders generated after Fourier transform. In one embodiment of this application, the second aperture stop 122 can filter out diffraction orders other than the +1st order.

[0066] To verify the real-time high-speed detection capability of the all-optical neural network system 100 of this application, Figure 3 The diagram shows the operational test results of an all-optical neural network system 100 according to an embodiment of this application. Figure 3 As shown, from Figure 3 Test results show that the real-time detection frame rate of the all-optical neural network system 100 of this application can reach 10752 fps (frames per second), currently limited by the overall switching rate of the digital micromirror device 102. When using a faster information source, the real-time detection frame rate of the all-optical neural network system 100 of this application can reach millions of frames per second, limited only by the bandwidth of the single-pixel high-speed detector 105.

[0067] It is understood that the optical path structure of the all-optical neural network system 100 described above is only an illustrative embodiment of this application; however, the optical path structure of the all-optical neural network system 100 is not limited to this embodiment. Figure 1 As shown, those skilled in the art can make several modifications and improvements without departing from the concept of this application, and all of them will be covered within the protection scope of this application.

[0068] This application also provides a method for optical information processing. Figure 4 A flowchart illustrating an embodiment of the optical information processing method of this application is provided. Figure 4 As shown, an embodiment of the optical information processing method of this application may include steps S1 to S5.

[0069] In step S1, the input image information is encoded onto the pulsed laser beam.

[0070] In step S2, a laser beam carrying the input image information is incident on a phase modulation device for phase modulation to output a light signal with a predetermined light intensity distribution, thereby completing the light speed inference of the input image.

[0071] In some embodiments, the phase modulation device of this application may include, for example, a spatial light modulator 103, which is used to construct a single-layer diffractive optical neural network structure and has a phase modulation function.

[0072] In step S3, different time delays are applied to the optical signals with predetermined light intensity distributions and then combined for output.

[0073] For example, an optical signal with a predetermined intensity distribution on the output plane of a phase modulation device can be coupled through a lens into different delay channels of an optical fiber delay line array device 104 for transmission and combined output.

[0074] In step S4, the merged output optical signal is coupled into a single-pixel high-speed photodetector for high-speed photoelectric conversion.

[0075] In step S5, the electrical signal output by the conversion is acquired and analyzed to obtain one-dimensional pulse time-domain waveform signal data, and then the input image information is obtained.

[0076] The optical information processing method of this application introduces a light field spatiotemporal projection method, which can be implemented based on the fiber delay line array device 104. The light field spatiotemporal projection method can be used to project the two-dimensional spatial light field into the one-dimensional time domain, thereby breaking through the slow reading speed limitation of the two-dimensional array camera.

[0077] The following will take a phase modulation device, including a spatial light modulator 103, as an example, combined with... Figure 5 This application will describe in detail the light field spatiotemporal projection method involved in a light information processing method according to an embodiment of the present application.

[0078] Figure 5 A schematic diagram illustrating a spatiotemporal projection method of the optical field based on a fiber delay line array device 104 according to an embodiment of this application is shown. Figure 5 As shown, one embodiment of the light field spatiotemporal projection method of this application includes: pre-optimizing the phase distribution of the spatial light modulator 103, and optimizing the two-dimensional light intensity distribution output by the neural network into multiple discretized spatial focal points, such as three spatial focal points. Then the first Each spatial focus is represented as The phase distribution is optimized using a constructed forward propagation algorithm and an error backpropagation algorithm, and the optimized phase distribution is then loaded onto the spatial light modulator 103. Three spatial focal points. Located on the output plane of the spatial light modulator 103.

[0079] The optical signals from each spatial focal point are incident on the input ends of corresponding optical fiber delay lines in the optical fiber delay line array device 104. Each optical fiber delay line has a different length, and different delay lines represent different optical signal delay channels. The number of optical fiber delay lines is the same as the number of spatial focal points. Furthermore, the arrangement of the multiple spatial focal points on the output plane of the spatial light modulator 103 is the same as the arrangement of the optical fiber delay lines in the optical fiber delay line array 1041 on the input plane of the optical fiber delay line array device 104. The output plane of the spatial light modulator 103 is also the input plane of the optical fiber delay line array device 104. A spatial focus The optical signal can be represented as It passes through a coupling lens and is incident on the input end of the fiber delay line array device 104, entering the corresponding... Fiber optic delay line.

[0080] The optical signals output from the individual fiber delay lines are combined into a time-domain multiplexed sequence by the combiner 1042 in the fiber delay line array device 104, and then output from a single fiber, completing the projection of the optical field from two-dimensional space to one-dimensional time. The time-domain multiplexed sequence can be converted into a high-speed electrical signal by the single-pixel high-speed detector 105, which is received by the signal acquisition and processing module 106 for input information reasoning and reconstruction.

[0081] Figure 6 The spatial light intensity distribution used in the light field spatiotemporal projection method of this application is revealed. For example... Figure 6 As shown, in one embodiment, the light intensity distribution output by the spatial light modulator 103 is optimized into seven light intensity distributions. Each light intensity distribution includes three spatial focal points. , , Arranged from left to right, if the sum of the light intensity at each spatial focal point exceeds a set detection threshold, it indicates that a light signal has been detected (represented by "1"); otherwise, it indicates that no light signal has been detected (represented by "0"). Therefore, the spatial encoding of the output light intensity distribution can be represented using 3 bits of data, from top to bottom as 001, 010, 011, 100, 110, 111. The arrangement of the spatial encoding in this embodiment is merely an illustrative implementation of this application; however, this application is not limited to this, and other arrangements of spatial encoding are also within the scope of protection of this application.

[0082] Optical signals in different fiber optic delay lines experience varying time delays during transmission. Among them, the first... The time delay of the optical signal corresponding to the fiber delay line It can be determined by its length Sure:

[0083]

[0084] in, Represents the speed of light. It is the group refractive index of the optical fiber, which is usually slightly larger than the optical fiber's refractive index.

[0085] In some embodiments, the gradient distribution of the lengths of each fiber delay line can be precisely controlled to generate a predetermined, resolvable time delay difference in the optical signal passing through each fiber delay line. :

[0086]

[0087] Among them, time delay difference Must meet ,in, The bandwidth of the single-pixel high-speed detector 105 is ensured to make the time-domain signal resolvable.

[0088] Temporal multiplexing sequence satisfy:

[0089]

[0090] in, Indicates the first The amplitude weights of each fiber delay line are used to adjust the signal strength of the channel; The shape of the laser pulse envelope; This indicates that the pulse was delayed. time; The center frequency of the optical carrier. It is the phase term of the carrier wave, and its relationship with time. and delay time Related. Indicates the first The phase noise term of each fiber delay line represents the phase jitter or instability of the signal. It is the imaginary unit.

[0091] The electrical signal can be acquired by the signal acquisition and processing module 106 at an ultra-high-speed sampling rate and analyzed to obtain one-dimensional pulse time-domain waveform signal data. The one-dimensional pulse time-domain waveform signal data can be used to extract the characteristic peaks of each fiber delay line through a matched filter bank to form a peak amplitude sequence. In one embodiment, the peak amplitude sequence [A1, A...] 2， A3] The time encoding that constitutes the output vector of the neural network compares each feature peak in the peak amplitude sequence with a set detection threshold to obtain the encoded information of the input image information.

[0092] In some embodiments, the spatiotemporal projection method of the light field in this application can introduce a delay compensation mechanism. A tunable delay line can be set at the output of the beam combiner 1042. The spatiotemporal projection method of the light field in this application can also include: using the tunable delay line to perform closed-loop feedback control on the output of the beam combiner 1042 to correct the time delay error caused by factors such as temperature drift and mechanical vibration in real time, with a compensation accuracy of ±0.5 ps.

[0093] Figure 7 This paper discloses a one-dimensional time-domain waveform signal obtained by a single-pixel high-speed detector 105 through high-speed switching of the input image on the digital micromirror device 102 using the light field spatiotemporal projection method of this application. Figure 7 As shown, each time-domain peak corresponds to a spatial focal point. By setting a detection threshold, 3-bit encoded information can be extracted from the pulse time-domain waveform signal. From top to bottom, the peak amplitude sequence is

[001] ,

[010] ,

[011] ,

[100] ,

[110] ,

[111] , which corresponds one-to-one with the spatial encoded information, thus verifying the feasibility of the light field spatiotemporal projection method proposed in this application.

[0094] The all-optical neural network system 100 and optical information processing method of one or more embodiments of this application have at least the following beneficial technical effects:

[0095] (1) Eliminate the limitations of two-dimensional detectors: adopt a time-series reading mechanism to avoid frame rate bottlenecks and support data reading at the GHz frequency level;

[0096] (2) Simple structure and low cost: The system only requires a single-pixel high-speed detector 105, which significantly reduces hardware complexity and cost;

[0097] (3) Strong system stability: The physical structure of the length difference of the fiber delay line is determined, the response consistency is high, and the anti-disturbance ability is strong;

[0098] (4) High scalability: The number and length difference of fiber delay lines can be freely adjusted according to the number of output channels to achieve different precision and resolution;

[0099] (5) Good compatibility: It is suitable for various optical neural network structures and has a universal high-speed output interface function.

[0100] In summary, the all-optical neural network system 100 and optical information processing method of this application can solve the problem of limited output readout speed in existing optical neural networks. The all-optical readout mechanism that uses optical fiber delay lines for space-time conversion, combined with a single-pixel high-speed detector 105, can significantly improve the actual inference speed of optical neural networks, expand the application scope of optical computing in the field of real-time artificial intelligence, and is suitable for information processing scenarios with high throughput and low latency.

[0101] The all-optical neural network system and optical information processing method provided in the embodiments of this application have been described in detail above. Specific examples have been used to illustrate the all-optical neural network system and optical information processing method of this application. The descriptions of the embodiments above are only for helping to understand the core ideas of this application and are not intended to limit this application. It should be noted that those skilled in the art can make several improvements and modifications to this application without departing from the spirit and principles of this application, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. An all-optical neural network system, characterized by: Comprising: a pulsed laser source for emitting a pulsed laser beam; a digital micromirror device for loading input image information and encoding the input image information onto the laser beam irradiated onto its surface; a phase modulation device for phase modulating the laser beam carrying the input image information out of the surface of the digital micromirror device to form a light signal of a predetermined light intensity distribution including a plurality of discrete spatial focal points on an output plane; a fiber delay line array device including a fiber delay line array and a beam combiner, the number of fiber delay lines in the fiber delay line array being the same as the number of spatial focal points, the fiber delay line array being used to apply different time delays to the light signal of the predetermined light intensity distribution, and the beam combiner being used to combine and output the light signals after applying different time delays; a single-pixel high-speed detector for receiving the light signal output from the fiber delay line array device and converting the light signal into an electrical signal at high speed; a signal acquisition and processing module for acquiring the electrical signal output from the single-pixel high-speed detector to obtain a one-dimensional pulsed time-domain waveform signal data, and further obtaining the input image information.

2. The system of claim 1, wherein: Further comprising a mirror for changing the path of the laser beam so that the laser beam is incident on the surface of the digital micromirror device at a predetermined angle.

3. The system of claim 1, wherein: Further comprising a 4f system for filtering the multi-order diffraction orders caused by the laser beam passing through the digital micromirror device, and only retaining the required diffraction orders to continue to propagate forward.

4. The system of claim 3, wherein: Further comprising a beam splitter for transmitting the light field out of the 4f system and reflecting the light field returned from the phase modulation device to the fiber delay line array device perpendicularly to the input light field.

5. The system of claim 1, wherein: Further comprising a beam expander and a first aperture stop, wherein, the beam expander is used to expand the diameter of the laser beam; the first aperture stop is used to filter the edges of the laser beam after expansion by the beam expander.

6. The system of claim 1, wherein: Further comprising a Fourier transform lens and a second aperture stop, wherein, the Fourier transform lens is used to apply Fourier transform to the output light field of the phase modulation device; the second aperture stop is used to filter the light intensity distribution of other diffraction orders generated after the Fourier transform.

7. The system of claim 1, wherein: The phase modulation device includes a spatial light modulator for constructing a single-layer diffractive optical neural network structure, wherein the phase distribution loaded on the spatial light modulator is obtained by pre-training on a computer using a deep learning algorithm to optimize the two-dimensional light intensity distribution output by the spatial light modulator into a plurality of discrete spatial focal points, wherein the arrangement of the plurality of spatial focal points on the output plane of the spatial light modulator is the same as the arrangement of the fiber delay lines in the fiber delay line array on the input plane of the fiber delay line array device, and each fiber delay line in the fiber delay line array has a different length.

8. The system of claim 7, wherein: The phase modulation device includes a plurality of spatial light modulators arranged at equal intervals according to the incident direction of the laser beam.

9. An optical information processing method, characterized by: Comprising: encoding input image information onto a pulsed laser beam; A laser beam carrying input image information is incident on a phase modulation device to perform phase modulation, so as to output a light signal with a predetermined light intensity distribution including a plurality of discrete spatial focal points; Different time delays are applied to the light signal with the predetermined light intensity distribution by a fiber delay line array device, and the light signal is output after being combined, the number of fiber delay lines in the fiber delay line array being the same as the number of spatial focal points; The combined output light signal is coupled into a single-pixel high-speed photodetector for high-speed photoelectric conversion; The converted electrical signal is collected and analyzed to obtain a one-dimensional pulse time-domain waveform signal data, and then the input image information is obtained.

10. The method of claim 9, wherein: The phase modulation device includes a spatial light modulator, and the method further comprises: The phase distribution of the spatial light modulator is optimized in advance on a computer by constructing a forward propagation algorithm and an error back propagation algorithm, and the optimized phase distribution is loaded onto the spatial light modulator, so as to optimize the two-dimensional light intensity distribution output by the spatial light modulator into a plurality of discrete spatial focal points.

11. The method of claim 10, wherein: The method of applying different time delays to the light signal with the predetermined light intensity distribution and combining and outputting the light signal comprises: The light signal of each spatial focal point is incident on the input end of each fiber delay line in the fiber delay line array device, respectively, wherein the lengths of the fiber delay lines are different, and the arrangement of the plurality of spatial focal points on the output plane of the spatial light modulator is the same as the arrangement of the fiber delay lines in the fiber delay line array on the input plane of the fiber delay line array device; The light signals output from the output ends of the fiber delay lines are combined into a time-domain multiplexing sequence by a combiner in the fiber delay line array device.

12. The method of claim 11, wherein: Further comprising: The gradient distribution of the length of each fiber delay line is controlled to cause a predetermined time delay difference between the light signals passing through each fiber delay line.

13. The method of claim 11, wherein: An adjustable delay line is arranged at the output end of the combiner, and the method further comprises: The output of the combiner is controlled by the adjustable delay line to perform closed-loop feedback control, so as to correct the time delay error in real time.

14. The method of claim 9, wherein: Further comprising: The one-dimensional pulse time-domain waveform signal data is extracted through a matched filter bank to form a peak amplitude sequence by extracting characteristic peaks passing through each fiber delay line; Each characteristic peak in the peak amplitude sequence is compared with a set detection threshold to obtain the encoding information of the input image information.

Images