Intelligent optical computing online training architecture and system

Abstract

The present invention discloses an intelligent optical computing online training architecture and system, which is applied to a free space system, including: using a solid-state laser to generate coherent light of a preset wavelength, using a beam expander to expand the coherent light wavefront, and using a beam splitter to divide the coherent light wavefront into a first path light beam and a second path light beam; inputting the first path light beam into a first spatial light modulator for data / error complex field loading, and using the second path light beam as an interference light beam; wherein the first spatial light modulator for field loading is configured to work in an amplitude modulation mode; relaying the amplitude modulated light field to the second spatial light modulator through a 4f system for phase modulation, adjusting the polarization with a half-wave plate, and passing the second path light beam through a half-wave plate and a linear polarizer to obtain an output light beam; using a detection symmetric propagation system to measure the output light beam to determine the amplitude and phase of the light beam measurement result. The present invention can avoid the complex reverse propagation process and improve the efficiency and stability of training.

Description

Technical Field

[0001] The present invention relates to the field of intelligent optical computing technology, and in particular to an intelligent optical computing online training architecture and system. Background Art

[0002] In an era of unprecedented advances in information technology, the pursuit of faster and more efficient computing methods has become a driving force for innovation. Optical computing is one of the unconventional computing paradigms that uses the unique properties of light to perform computing tasks, with the advantages of ultra-high processing speed, massive parallelism, and low power consumption. The appeal of optical computing lies in its use of the inherent properties of light, such as speed, parallelism, and coherence. Unlike traditional electronic computers that rely on the movement of electrons in semiconductor materials, optical computing uses photons (particles of light) to perform calculations. This paradigm shift opens up new possibilities for solving complex computing problems and is expected to revolutionize fields ranging from communications and image processing to artificial intelligence and pattern recognition.

[0003] In addition, in the field of artificial intelligence and machine learning, more efficient and scalable training methods have always been a pressing issue. Traditional electronic computing methods have problems such as slow speed, high energy consumption and poor scalability. Therefore, new alternative computing paradigms have begun to emerge, the most notable of which is optical training. Optical training represents an emerging paradigm in the field of machine learning that uses the unique properties of light to accelerate and enhance the training process. Unlike traditional electronic computing, optical training uses photons (particles of light) to perform computing tasks. This method has the advantages of ultra-high processing speed, huge parallelism and low energy consumption, so it has great potential in solving the increasingly complex challenges in the field of modern artificial intelligence and machine learning. Summary of the invention

[0004] The present invention aims to solve one of the technical problems in the related art at least to a certain extent.

[0005] To this end, the present invention proposes an online training method for intelligent optical computing, which is based on the technical pain point that current optical system calculations mainly rely on offline modeling and calculations, requiring a large amount of computing resources and accurate numerical models. Based on differentiable neural network parameters, the present invention performs optical training on a specific physical system and completes the training process of the optical network in an efficient way.

[0006] Another object of the present invention is to provide another online training method for intelligent optical computing.

[0007] The third objective of the present invention is to provide an intelligent optical computing online training system.

[0008] The fourth objective of the present invention is to provide another online training of intelligent optical computing.

[0009] To achieve the above object, the present invention provides an online training method for intelligent optical computing, which is applied to a free space system. The method comprises:

[0010] Generate coherent light of a preset wavelength using a solid-state laser, expand the coherent light wavefront using a beam expander, and split the coherent light wavefront into a first path light beam and a second path light beam using a beam splitter;

[0011] Inputting the first path light beam into the first spatial light modulator for data / error complex field loading, and using the second path light beam as an interference light beam; wherein the first spatial light modulator for field loading is configured to operate in an amplitude modulation mode;

[0012] The amplitude modulated light field is relayed to the second spatial light modulator through the 4f system for phase modulation, and after adjusting the polarization using a half-wave plate, the second path light beam is passed through the half-wave plate and the linear polarizer to obtain the output light beam;

[0013] The output beam is measured using a probe symmetric propagation system to determine the amplitude and phase of the beam measurement results.

[0014] The intelligent optical computing online training method of the embodiment of the present invention may also have the following additional technical features:

[0015] In one embodiment of the present invention, the amplitude modulation pattern is composed of 1280×1024 modulation elements, with a pitch of 12.5 μm and a programming depth of 8 bits.

[0016] In one embodiment of the present invention, the second spatial light modulator includes 1920×1200 modulation units, each modulation unit has a size of 8 μm and an 8-bit precision, the maximum frame rate of the second spatial light modulator is 60 Hz, and the update time is 17 m.

[0017] In one embodiment of the present invention, the detection symmetric propagation system includes at least a CMOS sensor, which is used to achieve a maximum frame rate of 4.2 Hz, corresponding to a pixel setting of 4096×2160, wherein the size of each pixel is 3.45 μm and the read value is configured as 8 bits.

[0018] To achieve the above object, the present invention further provides an online training method for intelligent optical computing, which is applied to an integrated photonic chip system. The method comprises:

[0019] A dual-channel input laser signal is generated by using a laser source;

[0020] Using a polarization controller to control the polarization of light in the optical fiber, and using a variable optical attenuator to shape and adjust the input laser signal after polarization control;

[0021] The adjusted input laser signal is optically coupled to a photonic chip through an input optical fiber array, and the dual-channel input laser signal is attenuated by a variable optical attenuator on the photonic chip, and is divided into two independent output paths to obtain output light;

[0022] The output light is passed through an output optical fiber array to generate a photocurrent, which is converted into a voltage signal after being amplified by a transimpedance amplifier. An oscilloscope is used to detect and analyze the voltage signal to obtain an analysis result.

[0023] To achieve the above-mentioned object, the present invention provides, on the other hand, an intelligent optical computing online training system, comprising: a light path splitting module, for generating coherent light of a preset wavelength using a solid-state laser, expanding the coherent light wavefront using a beam expander, and splitting the coherent light wavefront into a first path light beam and a second path light beam through a beam splitter;

[0024] A path beam loading module, used for inputting the first path beam into the first spatial light modulator for data / error complex field loading, and using the second path beam as an interference beam; wherein the first spatial light modulator for field loading is configured to work in an amplitude modulation mode;

[0025] The light field phase modulation module is used to relay the amplitude modulated light field to the second spatial light modulator through the 4f system for phase modulation, and after adjusting the polarization using the half-wave plate, the second path light beam is passed through the half-wave plate and the linear polarizer to obtain the output light beam;

[0026] The output result measurement module is used to measure the output light beam by using the detection symmetric propagation system to determine the amplitude and phase of the light beam measurement result.

[0027] To achieve the above object, the present invention further provides another intelligent optical computing online training system, comprising:

[0028] An input signal generating module, used for generating a dual-channel input laser signal by using a laser source;

[0029] An input signal adjustment module, for controlling the polarization of light in the optical fiber using a polarization controller, and shaping and adjusting the input laser signal using a variable optical attenuator after polarization control;

[0030] An output signal attenuation module, used for coupling the adjusted input laser signal to the photonic chip through the input optical fiber array, attenuating the dual-channel input laser signal on the photonic chip through a variable optical attenuator, and dividing it into two independent output paths to obtain output light;

[0031] The voltage signal analysis module is used to generate photocurrent through the output optical fiber array, convert the photocurrent into a voltage signal after amplification by a transimpedance amplifier, and use an oscilloscope to detect and analyze the voltage signal to obtain an analysis result.

[0032] The intelligent optical computing online training method and system of the embodiment of the present invention efficiently completes the training of optical networks, can train deep photonic neural networks with millions of parameters online, and can handle optical imaging systems in complex scenarios. The online training method greatly improves the training speed and achieves high-precision training results under resource-constrained conditions. Through the implementation of this architecture, it will usher in important applications in the fields of advanced machine learning, high-performance imaging, and topological photonics in the future.

[0033] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The above and / or additional aspects and advantages of the present invention will become apparent and easily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0035] Figure 1 is a flow chart of an online training method for intelligent optical computing according to an embodiment of the present invention;

[0036] Figure 2 is an architectural diagram of a free space system according to an embodiment of the present invention;

[0037] Figure 3 is a flow chart of another online training method for intelligent optical computing according to an embodiment of the present invention;

[0038] Figure 4 is an architecture diagram of an integrated photonic chip system according to an embodiment of the present invention;

[0039] Figure 5 is a schematic diagram of the structure of an intelligent optical computing online training system according to an embodiment of the present invention;

[0040] Figure 6 4 is a schematic diagram of the structure of another intelligent optical computing online training system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0041] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0042] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0043] The following describes the online training method and system for intelligent optical computing according to embodiments of the present invention with reference to the accompanying drawings.

[0044] Figure 1 It is a flow chart of the online training method of intelligent optical computing according to an embodiment of the present invention.

[0045] like Figure 1 As shown, applied to a free space system, the method comprises:

[0046] S1, using a solid-state laser to generate coherent light of a preset wavelength, using a beam expander to expand the coherent light wavefront, and using a beam splitter to split the coherent light wavefront into a first path light beam and a second path light beam;

[0047] S2, inputting the first path light beam into the first spatial light modulator for data / error complex field loading, and using the second path light beam as an interference light beam; wherein the first spatial light modulator for field loading is configured to operate in amplitude modulation mode;

[0048] S3, relaying the amplitude modulated light field to the second spatial light modulator through the 4f system for phase modulation, adjusting the polarization using a half-wave plate, and then passing the second path light beam through the half-wave plate and the linear polarizer to obtain an output light beam;

[0049] S4, measuring the output beam using a detection symmetric propagation system to determine the amplitude and phase of the beam measurement results.

[0050] It can be understood that the embodiments of the present invention realize efficient and high-precision training of optical systems of large-scale intelligent optical computing systems. The premise for realizing optical training is to divide the system into a modulation region and a propagation region. In the modulation region, the refractive index can realize various functions through electro-optical or all-optical reconstruction, while in the propagation region, the refractive index remains fixed. This division provides a basic framework for optical training, so that the parameters of the optical system can be flexibly adjusted and optimized. Full forward mode gradient descent learning is one of the key methods for optical training. Unlike traditional back-propagation-based methods, full forward mode gradient descent learning does not require back-propagation of the light field, but calculates the gradient relative to the refractive index by feeding the error field into the system and measuring the light field at the output. The advantage of this method is that it can avoid the complex back-propagation process and improve the efficiency and stability of training.

[0051] It is understood that the basic elements of the free space optical system of the embodiment of the present invention are refractive index, gain and loss, which can be summarized as the real part and imaginary part of the refractive index: n = n R +in I Here, the present invention will prove two conclusions. First, the linear propagation of the optical system can be restated as a linearly connected neural network. Second, the neural network is differentiable and the system parameters can be designed by gradient descent.

[0052] Consider an optical system governed by Maxwell's equations,

[0053] Where E is the electric field distribution, J is the current source, and ∈0,∈ r ,μ0,μ r are vacuum, relative permittivity and magnetic permeability respectively. Maxwell's equations can be vectorized as A(∈ r )E=μ r J. Since the situation under consideration is caused by the same current source J, input electric field E i and the output electric field E o These all satisfy A i E i =μ r J and A o E o =μ r J. Then the inventors began to prove that in a linear system, there exists a pseudo-inverse matrix Make Here I is the identity matrix.

[0054] Proof: Let and i is a one-hot vector with non-zero position i, then j i Will be the base current source J = Σ i α i j i For each j i As the source of the system, the output electric field distribution is expressed as A o e i = i Since the system is linear, the output This shows Compare this to the governing forward Maxwell equation A o E o =J, the conclusion of the present invention is satisfy

[0055] Will Substituting into Maxwell's equations for input and output fields, the input field E i and output field E o It can be connected by the following equation:

[0056]

[0057] here and i Parameterized as ∈ r The present invention can be observed from the above that the linear propagation representing the light field depicts a linear layer connecting the input and output with a fully connected weight matrix The present invention further considers the gradient descent design of parameters and the relay between layers. First, it is observed that there are two forms of wave field manipulation: modulation and coherent interference. Modulation can be expressed as point-by-point multiplication, and coherent interference can be expressed as addition in complex fields. Consider the modulation process: Where Δd is the modulation distance, λ is the wavelength, n R ,n I are the real and imaginary parts of the refractive index. When combining coherent reference beams, the present invention has So the input field expands to The end-to-end embedded neural network representation is

[0058]

[0059] here and Maps to the inputs, outputs, complex multiplications, and bias weights in a neural network implementation. Here and denote the refractive index of the fixed region and the adjustable region, respectively. Please note that for simplicity, the reference input is omitted in the main text of this invention. In practical applications, the system is designed for a target T, and the related quantitative loss function is denoted as L. The gradient of the complex refractive index is for

[0060]

[0061] In addition, the gradients of the input system and the reference system are

[0062]

[0063] In this way, the gradients of all parameters in the system can be retrieved and designed via gradient descent.

[0064] For a multi-layer parameterized optical system, the present invention will introduce a gradient descent algorithm for the system. The propagation matrix W of the kth linear optical layer in the above expression is re-expressed as the Green function G k (ro ,r i ), then the forward propagation can be expressed as:

[0065]

[0066] The output can be nonlinearly activated and fed into subsequent layers x k+1 =f(y k ). The quality factor is evaluated based on the final output, L = Ψ (y N ,T) where T is the target output. To effectively evaluate the loss, the gradient is calculated using the loss function L. The gradient y N for:

[0067]

[0068] According to the chain rule, the gradient propagation is:

[0069]

[0070] According to Lorentz reciprocity, G k (r o ,r i )=G k (r i r o ), so that

[0071]

[0072] This corresponds to the propagation field δ y From the output of layer k to the input of layer k.

[0073] In the embodiment of the present invention, an online training method is applied to a free space system to conduct experiments, such as Figure 2 shown.

[0074] Specifically, in the free-space system, the source of optical computing is generated using a solid-state laser (MFL-FN-532) with an operating wavelength of 532nm. The coherent light wavefront is expanded using a beam expander and then split into two paths by a beam splitter (BS013, Thorlabs). One path is directed to a spatial light modulator (SLM) for data / error complex field loading, while the other path is used as an interference beam, and its light intensity is attenuated by two polarizers (LPNIR100-MP2, Thorlabs). The SLM (X15213-01, HAMAMATSU) used for field loading is configured to work in amplitude modulation mode, which consists of 1280×1024 modulation elements with a pitch of 12.5μm and a programming depth of 8 bits. It can modulate the incident light at a speed of up to 60Hz. The amplitude modulated light field is then relayed 1:1 through a 4f system to another SLM (E series, meadowlark) for phase modulation, after adjusting the polarization with a half-wave plate. Then it passes through a half-wave plate and a linear polarizer. The phase SLM is a reflective silicon-based liquid crystal device with a zero-order diffraction efficiency of up to 91%. It contains 1920×1200 modulation units, each with a size of 8μm and 8-bit accuracy. The maximum frame rate of the phase SLM is 60Hz, and the update time is 17ms. The phase modulation SLM is suitable for both electric field loading and designed phase loading. The wavefront carries the input light field and the designed phase, and then detects the symmetrical propagation system to perform a specific task. The output beam is split into two independent sensors before propagating a length of 0.388m. A CMOS sensor (BFS-U3-89S6M, Flir) is used to measure complex fields, good quantum efficiency (63.99%) and low dark noise (2.47e-). Two sensors are used to measure the amplitude and phase of the results respectively. The CMOS sensor can achieve a maximum frame rate of 4 2Hz, with a pixel setting of 4096×2160, where the size of each pixel is 3.45μm and the readout value is configured as 8 bits. For a free-space system, each propagation takes the same amount of time, regardless of system complexity. Commercial spatial light modulators can exceed 1kHz (HSP1K-488-800-PC8, Meadowlark Optics), so the propagation time for each iteration is less than 2.0ms.

[0075] The online training method for intelligent optical computing of the embodiment of the invention can avoid the complex back-propagation process and improve the efficiency and stability of training. The optical training is performed on a specific physical system, and the training process of the optical network is completed in an efficient way. The system is capable of online training of deep photonic neural networks with millions of parameters, as well as processing optical imaging systems in complex scenarios. The online training method greatly improves the training speed and achieves high-precision training results under resource-constrained conditions.

[0076] Figure 3 The intelligent optical computing online training method of the embodiment of the present invention is applied to an integrated photonic chip system, and the method includes:

[0077] S10, generating a dual-channel input laser signal using a laser source;

[0078] S20, controlling the polarization of light in the optical fiber using a polarization controller, and shaping and adjusting the input laser signal using a variable optical attenuator after polarization control;

[0079] S30, coupling the adjusted input laser signal to the photonic chip through the optical coupling of the input optical fiber array, attenuating the dual-channel input laser signal on the photonic chip through a variable optical attenuator, and dividing it into two independent output paths to obtain output light;

[0080] S40, the output light is passed through the output optical fiber array to generate a photocurrent, which is converted into a voltage signal after being amplified by a transimpedance amplifier, and the voltage signal is detected and analyzed by an oscilloscope to obtain an analysis result.

[0081] Specifically, if Figure 4 As shown, in an integrated photonic chip system according to an embodiment of the present invention, the chip is produced in a silicon photonic foundry. The chip integrates various components such as passive grating couplers, waveguides and multi-mode interferometers, which are manufactured on a silicon on insulator (SOI) platform using a standard etching process.

[0082] Specifically, two laser sources with a power of 10 mW are used in the system to generate dual-channel input. One laser operates at a precise wavelength of 1550 nm, while the other laser operates at a precise wavelength of 1551 nm. A polarization controller is used to control the polarization of light within the fiber to ensure that the desired polarization state is achieved. After polarization control, a variable optical attenuator (V1550A, Thorlabs) is used to shape and adjust the input signal. The output of the variable optical attenuator then passes through the input fiber array to facilitate the coupling of light to the photonic chip. On the chip, four variable optical attenuators (VOAs) attenuate the dual-channel input light and separate it into two independent output paths. Subsequently, the attenuated light is detected by an on-chip photodiode (DET01CFC, Thorlabs). The output light passes through the output fiber array, and the generated photocurrent is amplified by a transimpedance amplifier (TIA, AMP100, Thorlabs) and converted into a voltage signal. Finally, an oscilloscope is used to detect and analyze the dual-channel output. In addition, an eight-channel fiber array with an angle of 8° is used to guide light in and out of the photonic chip. After aligning the fiber array with the grating array under a microscope with the help of a nano-motorized translation stage, the vertically coupled fiber array was connected and fixed to the PIC with a curable epoxy resin. The insertion loss of the fiber-to-chip coupling measured after packaging was about -8dB per channel. Two layers of dedicated PCB boards were used for on-chip electrical signal feeding. The VOA array on-chip pads with a period of 100μm were connected to the PCB board through gold wire bonding and independently routed to the electrical socket through signal lines with a period of 800μm. A customized multi-channel DC signal source was connected to the board, and the injection current of the VOA was controlled by setting the voltage 0-5V, which effectively adjusted the imaginary part of the effective refraction of the guided mode wave. For thermal stability, the chip core and thermistor were firmly mounted on the copper block using thermal adhesive. The thermistor measured the temperature of the chip, and a thermoelectric temperature controller (TEC, Peltier cooler) was connected to the copper block to cool the packaging system. A proportional integral derivative (PID) feedback loop was established between the thermistor and the TEC. This feedback loop is able to maintain the on-chip temperature within ±0.004 Kelvin (equivalent to ±2Ω thermistor long-term stability), ensuring reliable and consistent operation.

[0083] It is known that most of the training operations are implemented experimentally online in different optical systems. Here, the invention explains how to perform the remaining operations on site. Both loss calculations and gradient calculations can be implemented by on-site optical or electronic devices. The loss calculation involves a minus operation, which can be implemented by a π phase shift (equivalent to a -1 value in the complex representation of light), and an interference operation, calculating the difference T between the output y and the target, e = yT.

[0084] The gradient of the parameter g∝(Re{1jE data E error}-Im{1jEdata E error}). Therefore, the gradient calculation involves calculating the real and imaginary parts of the field E data E error In order to calculate Re{1jE data E error}, the strength of three fields needs to be detected: B=|E data | 2 、C=|E error | 2 , the gradient sum is Re{1jE data E error}=(ABC) / 2, where phase conjugation can be achieved by phase conjugation mirror, and differential operation can be achieved by balanced photoelectric detection. can be measured, then the present invention has Im{1jE data E error}=(A ′ -BC) / 2. The detected gradient signal will be further accumulated by an integration circuit to determine the modulator and tune the refractive index. The gradient can also be used to directly update the refractive index of the non-volatile optical material.

[0085] The online training method for intelligent optical computing of the present invention is inevitably imperfect due to the noise and general uncertainty that are prevalent in the physical nature. The online machine learning method proposed by the present invention alleviates the limitations of numerical modeling, realizes high-performance general optical artificial intelligence and optical artificial intelligence, and experimentally proves the online training method of the integrated photonic chip system, performs optical training on a specific physical system, and completes the training process of the optical network in an efficient way.

[0086] In order to implement the above embodiment, Figure 5 As shown, this embodiment also provides an intelligent optical computing online training system 10, including:

[0087] A light path beam splitting module 100, for generating coherent light of a preset wavelength using a solid-state laser, expanding the coherent light wavefront using a beam expander, and splitting the coherent light wavefront into a first path light beam and a second path light beam through a beam splitter;

[0088] A path beam loading module 200 is used to input the first path beam into a first spatial light modulator for data / error complex field loading, and use the second path beam as an interference beam; wherein the first spatial light modulator for field loading is configured to work in an amplitude modulation mode;

[0089] The light field phase modulation module 300 is used to relay the amplitude modulated light field to the second spatial light modulator through the 4f system for phase modulation, and after adjusting the polarization using a half-wave plate, pass the second path light beam through the half-wave plate and the linear polarizer to obtain an output light beam;

[0090] The output result measurement module 400 is used to measure the output light beam using a detection symmetric propagation system to determine the amplitude and phase of the light beam measurement result.

[0091] According to the intelligent optical computing online training system of the embodiment of the invention, the complex back-propagation process can be avoided, and the efficiency and stability of training can be improved. The optical training is performed on a specific physical system, and the training process of the optical network is completed in an efficient way. The system can train deep photonic neural networks with millions of parameters online, and process optical imaging systems in complex scenes. The online training method greatly improves the training speed and achieves high-precision training results under resource-constrained conditions.

[0092] In order to implement the above embodiment, Figure 6 As shown, this embodiment also provides another intelligent optical computing online training system 20, including:

[0093] An input signal generating module 30, used to generate a dual-channel input laser signal using a laser source;

[0094] An input signal adjustment module 40, for controlling the polarization of light in the optical fiber using a polarization controller, and shaping and adjusting the input laser signal using a variable optical attenuator after polarization control;

[0095] An output signal attenuation module 50 is used to couple the adjusted input laser signal to the photonic chip through the input optical fiber array, attenuate the dual-channel input laser signal on the photonic chip through a variable optical attenuator, and divide it into two independent output paths to obtain output light;

[0096] The voltage signal analysis module 60 is used to generate a photocurrent through the output optical fiber array, convert the photocurrent into a voltage signal after amplification by a transimpedance amplifier, and detect and analyze the voltage signal using an oscilloscope to obtain an analysis result.

[0097] The intelligent optical computing online training system of the present invention is inevitably imperfect due to the noise and general uncertainty that are ubiquitous in the physical nature. The online machine learning method proposed by the present invention alleviates the limitations of numerical modeling, realizes high-performance general optical artificial intelligence and optical artificial intelligence, and experimentally proves the online training of the integrated photonic chip system, performs optical training on a specific physical system, and completes the training process of the optical network in an efficient way.

[0098] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0099] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In the description of the present invention, the meaning of "plurality" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

Claims

1. An online training method for intelligent optical computing, characterized in that: Applied to a free space system, the method comprises: Generate coherent light of a preset wavelength using a solid-state laser, expand the coherent light wavefront using a beam expander, and split the coherent light wavefront into a first path light beam and a second path light beam using a beam splitter; Inputting the first path light beam into the first spatial light modulator for data / error complex field loading, and using the second path light beam as an interference light beam; wherein the first spatial light modulator for field loading is configured to operate in an amplitude modulation mode; The amplitude modulated light field is relayed to the second spatial light modulator through the 4f system for phase modulation, and after adjusting the polarization using a half-wave plate, the second path light beam is passed through the half-wave plate and the linear polarizer to obtain the output light beam; The output beam is measured using a probe symmetric propagation system to determine the amplitude and phase of the beam measurement results.

2. The method according to claim 1, characterized in that The amplitude modulation pattern consists of 1280×1024 modulation elements, with a pitch of 12.5 μm and a programming depth of 8 bits.

3. The method according to claim 1, characterized in that The second spatial light modulator includes 1920×1200 modulation units, each modulation unit has a size of 8 μm and an 8-bit precision, a maximum frame rate of the second spatial light modulator is 60 Hz, and an update time is 17 ms.

4. The method according to claim 1, characterized in that The detection symmetric propagation system at least includes a CMOS sensor, which is used to achieve a maximum frame rate of 4.2 Hz, corresponding to a pixel setting of 4096×2160, wherein the size of each pixel is 3.45 μm and the read value is configured as 8 bits.

5. An online training method for intelligent optical computing, characterized in that: Applied to an integrated photonic chip system, the method comprises: A dual-channel input laser signal is generated by using a laser source; Using a polarization controller to control the polarization of light in the optical fiber, and using a variable optical attenuator to shape and adjust the input laser signal after polarization control; The adjusted input laser signal is optically coupled to a photonic chip through an input optical fiber array, and a variable optical attenuator is used on the photonic chip to attenuate the dual-channel input laser signal and divide it into two independent output paths to obtain output light; The output light generates a photocurrent through an output optical fiber array, which is converted into a voltage signal after being amplified by a transimpedance amplifier, and an oscilloscope is used to detect and analyze the voltage signal to obtain an analysis result; The method further comprises: After aligning the input fiber array and the output fiber array with the grating array respectively, the vertically coupled fiber array is connected and fixed to the integrated circuit of the PCB board using a curable epoxy resin; and, The electrical signal on the photonic chip is fed through two layers of PCB boards, and the pads on the variable optical attenuator array chip with a period of 100 μm are connected to the PCB board through gold wire bonding, and are independently wired to the electrical socket through signal lines with a period of 800 μm; The method further comprises: Connect a multi-channel DC signal source to the PCB board, and control the injection current of the variable optical attenuator by setting the voltage 0-5V to adjust the imaginary part of the effective refraction of the guided mode wave; The method further comprises: The photonic chip core and the thermistor are mounted on a copper block using a thermal adhesive, the temperature of the chip is measured using the thermistor, a thermoelectric temperature controller is connected to the copper block to cool the packaging system, and a proportional-integral-differential feedback loop is established between the thermistor and the thermoelectric temperature controller; The method further includes: calculating the loss and calculating the gradient, wherein: The loss calculation involves a subtraction operation, achieved by a π phase shift, and an interference operation, calculating the difference between the output y and the target, T, e = yT; The gradient of the parameter g∝(Re{1jE data E error }-Im{1jE data E error }); Gradient calculation involves calculating the real and imaginary parts of the field E data E error ; Calculate Re{1jE data E error }, the strength of three fields needs to be detected: B=|E data | 2 、C=|E error | 2 , the gradient sum is Re{1jE data E error }=(ABC) / 2, where phase conjugation is achieved through a phase conjugation mirror and differential operation is achieved through balanced photodetection; intensity Being measured, there is Im{1jE data E error }=(A ′ -BC) / 2; the detected gradient signal is accumulated through an integration circuit to determine the modulator and tune the refractive index.

6. An intelligent optical computing online training system, characterized in that: include: A light path beam splitting module, used to generate coherent light of a preset wavelength using a solid-state laser, expand the coherent light wavefront using a beam expander, and split the coherent light wavefront into a first path light beam and a second path light beam through a beam splitter; A path beam loading module, used for inputting the first path beam into the first spatial light modulator for data / error complex field loading, and using the second path beam as an interference beam; wherein the first spatial light modulator for field loading is configured to work in an amplitude modulation mode; The light field phase modulation module is used to relay the amplitude modulated light field to the second spatial light modulator through the 4f system for phase modulation, and after adjusting the polarization using the half-wave plate, the second path light beam is passed through the half-wave plate and the linear polarizer to obtain the output light beam; The output result measurement module is used to measure the output light beam by using the detection symmetric propagation system to determine the amplitude and phase of the light beam measurement result.

7. An intelligent optical computing online training system using the method as claimed in claim 5, characterized in that: include: An input signal generating module, used for generating a dual-channel input laser signal by using a laser source; An input signal adjustment module, for controlling the polarization of light in the optical fiber using a polarization controller, and shaping and adjusting the input laser signal using a variable optical attenuator after polarization control; An output signal attenuation module, used to couple the adjusted input laser signal to the photonic chip through an input optical fiber array, attenuate the dual-channel input laser signal on the photonic chip through a variable optical attenuator, and divide it into two independent output paths to obtain output light; The voltage signal analysis module is used to generate photocurrent through the output optical fiber array, convert the photocurrent into a voltage signal after amplification by a transimpedance amplifier, and use an oscilloscope to detect and analyze the voltage signal to obtain an analysis result.

Images