Training method of coherent ising machine and related equipment

Abstract

The application relates to a training method of a coherent Ising machine and related equipment. The training method comprises the following steps: injecting noise into the coherent Ising machine, so that, under the joint action of physical parameters and the noise, the continuous amplitude vector of the light pulse generated by the coherent Ising machine evolves to a steady-state distribution state subject to the Gibbs distribution according to the Langevin dynamics evolution rule, the physical parameters include pumping parameters, saturation parameters, coupling coefficients, bias field parameters and gains; after the continuous amplitude vector collected under the current steady-state distribution is taken as negative-phase amplitude data, one or more of the physical parameters are updated based on the update amount obtained from the deviation between the negative-phase amplitude data and positive-phase amplitude data, so that the continuous amplitude vector reaches a target steady-state distribution. The application solves the problem that in the traditional CIM machine, the parameters are mostly fixed or empirically adjusted, and the steady-state sample distribution of the light pulse changes greatly due to environmental drift and device errors.

Description

Technical Field

[0001] This application relates to the field of quantum computing technology, and in particular to a training method and related equipment for a coherent Ising machine. Background Technology

[0002] In recent years, combinatorial optimization, probabilistic sampling, and generative modeling have been widely applied in fields such as communication resource allocation, chip design, image processing, anomaly detection, industrial control, and intelligent manufacturing. Traditional general-purpose computers (CPU and GPU) often face problems such as high power consumption, high latency, high heat dissipation pressure, communication bottlenecks, and limited scalability when processing the search and sampling of large-scale non-convex energy landscapes.

[0003] Specialized hardware based on nonlinear oscillator networks (such as optical parametric oscillator networks) leverages the parallel evolution characteristics of physical systems to demonstrate potential advantages over some problems encountered by traditional general-purpose computers. For example, coherent Ising machines (CIM) hardware based on DOPO networks (i.e., degenerate optical parametric oscillator networks): early research proposed using degenerate optical parametric oscillator networks to solve NP-hard problems in the Ising form of the Ising model (i.e., stochastic process models used to describe phase transitions of matter). Typical schemes construct coupled networks through optical delay lines or measurement feedback.

[0004] While existing CIM / DOPO network hardware has shown potential in some optimization tasks, it still has at least the following shortcomings when facing more general scenarios such as "controllable sampling of probability distributions, generative modeling, and steady-state distribution fitting": Uncontrollable or difficult to accurately calibrate steady-state distribution: Traditional CIM often aims to "find the optimal solution / low energy state". Hardware parameters (pump, saturation, gain, coupling) are mostly fixed or empirically tuned, which causes the steady-state sample distribution of optical pulses to change significantly with environmental drift and device errors, making it difficult to achieve stable reproduction across different tasks. Summary of the Invention

[0005] The main objective of this application is to provide a training method and related equipment for a coherent Ising machine, so as to at least solve the problem mentioned in the background art that the parameters of traditional CIM machines are mostly fixed or empirically tuned, resulting in significant changes in the steady-state sample distribution of optical pulses due to environmental drift and device errors.

[0006] According to one aspect of this application, a training method for a coherent Ising machine is provided, the coherent Ising machine including a laser pulse generator for generating N optical pulses with continuous amplitudes, where N is a natural number, and the N continuous amplitudes constitute an N-dimensional continuous amplitude vector. The training method includes: Noise is injected into the coherent Ising machine so that, under the combined action of the physical parameters of the coherent Ising machine and the noise, the continuous amplitude vector evolves to a steady-state distribution that follows the Gibbs distribution according to the Langevin dynamics evolution rule. The physical parameters include pump parameters, saturation parameters, coupling coefficients, bias field, and gain. After taking the continuous amplitude vector collected under the current steady-state distribution as negative phase amplitude data, one or more of the physical parameters are updated based on the update amount obtained from the deviation between the negative phase amplitude data and the positive phase amplitude data, so that the continuous amplitude vector reaches the target steady-state distribution. The positive phase amplitude data comes from a pre-set training dataset.

[0007] Further, the step of obtaining the update amount based on the deviation between the negative phase amplitude data and the positive phase amplitude data includes: After calculating the first statistic of the negative phase amplitude data and the second statistic of the positive phase amplitude data respectively, the update quantity is generated based on the difference between the first statistic and the second statistic, wherein the first statistic and the second statistic each include one or more of the following statistics: The first moment, which is a statistic required to update the bias field, represents the amplitude of the i-th optical pulse. The mean of i∈N; The second moment, which is a statistic required to update the pump parameters, represents the amplitude of the i-th optical pulse. The average of the squares; The second-order correlation moment, which is a statistic required to update the coupling coefficient, represents the amplitude of the i-th optical pulse. and the amplitude of the j-th optical pulse The average of the products between them, j∈N; The fourth moment, which is a statistic required to update the saturation parameter, represents the average value of the fourth power of the amplitude of the i-th light pulse; The amplitude energy correlation term is a statistic required to update the gain. The amplitude energy correlation term is the average energy of the coupling term and the bias term. The coupling term is the total energy of the interaction between all pulse pairs formed by the i-th optical pulse and the j-th optical pulse. The bias term is the energy of the i-th optical pulse under the action of the bias field.

[0008] Furthermore, the update amount includes an update amount used to update the bias field amount. Then, the step of generating the update quantity based on the difference between the first statistic and the second statistic includes: Obtain the first moment of the first statistic The first moment of the second statistic The intensity D of the noise and the learning rate required to update the bias field quantity. ; In calculation and After calculating the difference between them, the obtained difference is compared with the gain and the learning rate. The product of the three factors, including the reciprocal of D, is added to the bias field before the update to obtain the update amount. .

[0009] Furthermore, the update amount also includes an update amount for updating the pump parameters. Then, the step of generating the update quantity based on the difference between the first statistic and the second statistic further includes: Obtain the second moment of the first statistic The second moment of the second statistic The noise intensity D and the learning rate required to update the pump parameters. ; In calculation After calculating the difference between them, the obtained difference is then compared with the learning rate. The product of the inverse of 2D and the pump parameter before the update is added to obtain the update amount. .

[0010] Furthermore, the update amount also includes an update amount used to update the coupling coefficient. Then, the step of generating the update quantity based on the difference between the first statistic and the second statistic further includes: Obtain the second-order correlation moments of the first statistic The second correlation moments of the second statistic The noise intensity D and the learning rate required to update the coupling coefficients. ; In calculation and After calculating the difference between them, the obtained difference is compared with the learning rate. The product of the inverse of 2D and the gain is added to the coupling coefficient before the update to obtain the update amount. .

[0011] Furthermore, the method also includes: The coupling coefficient between all the pulse pairs is set. The coupling matrix J formed is symmetric and When =0, based on the update amount After updating the coupling coefficient, a symmetric projection operation and a zero diagonal constraint operation are performed on the coupling coefficient. The symmetric projection operation is the operation of projecting the coupling matrix J onto a symmetric matrix space, and the zero diagonal constraint operation is the operation of setting the elements on the diagonal of the coupling matrix J to zero.

[0012] Furthermore, the update amount also includes an update amount for updating the saturation parameter. Then, the step of generating the update quantity based on the difference between the first statistic and the second statistic further includes: Obtain the fourth moment of the first statistic The fourth moment of the second statistic The noise intensity D and the learning rate required to update the saturation parameter. ; In calculation and After calculating the difference between them, the obtained difference is compared with the learning rate. The product of the inverse of 4D and the saturation parameter before the update is added to obtain the update amount. .

[0013] Furthermore, the update amount also includes an update amount for updating the gain. Then, the step of generating the update quantity based on the difference between the first statistic and the second statistic further includes: Obtain the amplitude energy related term of the first statistic. The amplitude energy related term of the second statistic The intensity D of the noise and the learning rate required to update the gain. ,in, Let be the coupling coefficient between the i-th optical pulse and the j-th optical pulse. For coupling terms, For bias terms; In calculation and After calculating the difference between them, the obtained difference is compared with the reciprocal of D and the learning rate. The product of the two is added to the gain before the update to obtain the update amount. .

[0014] Furthermore, the method also includes: Through the update amount During the update of the gain, the magnitude of the gain change is constrained based on set constraint rules. These constraint rules include magnitude constraints or regularization terms. The magnitude constraints limit the range of values ​​for the gain, and the regularization terms are used to calculate the update amount. A penalty term is added during the process to suppress excessive changes in the gain.

[0015] Further, the step of updating the physical parameters based on the update amount includes: The update amount used for the pump parameters The update is sent to the pump modulation unit of the coherent Ising machine so that the pump modulation unit can adjust the update based on the update amount. Modulate the optical pulse; and / or, The update amount used to update the saturation parameter The update is sent to the saturation control unit of the coherent Ising machine, so that the saturation control unit can update the amount of data. Control the saturation parameter within a stable range; and / or, The update amount used to update the gain Update the amount of the coupling coefficient. And the update amount of the bias field quantity. The data is sent to the feedback module of the coherent Ising machine so that the feedback module can update the data based on the continuous amplitude under the current steady-state distribution of the optical pulse and the update amount. The update amount and the update amount The obtained feedback quantity is used to adjust the optical pulse, and the feedback quantity includes the total amount of the coupling effect of each optical pulse on other optical pulses and the amount of the bias field it receives.

[0016] Furthermore, the method also includes: The negative phase amplitude data and the positive phase amplitude data are input to the calculation unit of the feedback module to calculate the first statistic and the second statistic, and the calculation unit generates the update quantity based on the difference between the first statistic and the second statistic.

[0017] According to another aspect of this application, a computing device based on a coherent Ising machine is provided, the computing device being used to execute the training method of the coherent Ising machine, the device comprising: A coherent Ising machine includes a laser pulse generator, a control module, and a feedback module. The control module and the feedback module are respectively connected to the laser pulse generator. The laser pulse generator is used to generate N optical pulses with continuous amplitude. The control module is used to control the pump parameters and saturation parameters of the optical pulses. The feedback module is used to adjust the optical pulses based on feedback values ​​obtained from the continuous amplitude, gain, coupling coefficient, and bias field of the optical pulses. A noise controllable injector is provided, which is connected to the laser pulse generator and the feedback module respectively. The noise controllable injector is used to generate controllable noise and inject the noise into the laser pulse generator or the feedback module. A training controller is connected to the control module and the feedback module. The training controller is used to obtain an update amount based on the deviation between the collected negative phase amplitude data and positive phase amplitude data, and to send the update amount to the control module and the feedback module.

[0018] Furthermore, the update amount includes the update quantity. Update volume Update volume Update volume and update volume The control module includes: A pump modulation unit, wherein the pump modulation unit is used to modulate the pump parameters; A saturation control unit is used to control the saturation parameter within a stable range. Both the saturation control unit and the pump modulation unit are connected to the training controller. The training controller will update the amount. The update amount is sent to the pump modulation unit. The update amount is sent to the saturation control unit. The update amount and the update amount The data is then sent to the feedback module.

[0019] Furthermore, the coherent Ising machine also includes a measurement unit for measuring the continuous amplitude of the optical pulse, and the feedback module includes: A calculation unit is connected to both the measurement unit and the training controller, and the calculation unit generates the feedback quantity after receiving the continuous amplitude. A feedback injection unit is connected to the calculation unit and the laser pulse generator. The feedback injection unit is used to inject the feedback amount into the laser pulse generator to perform feedback adjustment on the optical pulse.

[0020] The training controller can calculate a first statistic and a second statistic through the computing unit, and generate the update quantity based on the difference between the first statistic and the second statistic through the computing unit.

[0021] Furthermore, the feedback module also includes: A sampling unit is connected between the measurement unit and the computing unit. The sampling unit is used to send amplitude data collected from the continuous amplitude to the computing unit during the evolution of the coherent Ising machine under the current physical parameters and noise.

[0022] Furthermore, the computing unit includes: A control chip is connected to both the measurement unit and the training controller. The control chip includes either a field-programmable gate array (FPGA) or an application-specific integrated circuit (ASIC). A computing device connected to the control chip, the computing device comprising either a multiplication and accumulation MAC array or a cross array.

[0023] According to another aspect of this application, a non-transitory machine-readable medium storing computer instructions for causing the computer to perform the method is provided.

[0024] The beneficial effects of the embodiments of this application are as follows: The training method for the coherent Ising machine provided in this application, after injecting noise into the coherent Ising machine, determines the continuous amplitude of the output light pulse by both its physical parameters and the noise. The change (or magnitude) of the continuous amplitude of the light pulse is equivalent to the Brownian motion of a tiny particle in Langevin dynamics. The pump parameters, saturation parameters, coupling coefficient, bias field, and gain are equivalent to deterministic forces applied to the light pulse, while the noise is a random force applied to the light pulse. This allows the coherent Ising machine to be modeled as an optical Langevin machine, whose output light pulse state is always a continuous optical amplitude variable. Therefore, the continuous amplitude of the N light pulses output by the coherent Ising machine will evolve to a steady-state distribution following a Gibbs distribution, thus enabling the coherent Ising machine to be considered a physically realized energy model sampler.

[0025] Therefore, the coherent Ising machine obtained in this application can be trained like an energy model, and the energy parameters for training such a coherent Ising machine can be fully physicalized into its physical parameters. Based on this, after using the continuous amplitude vector collected under the current steady-state distribution as negative phase amplitude data, this application can update the physical parameters based on the update amount obtained from the deviation between the negative and positive phase amplitude data, so that the steady-state distribution reaches the target steady-state distribution. Thus, during the operation of the coherent Ising machine, the physical parameters can be updated in real time according to the collected continuous amplitude, causing the amplitude of the light pulse to gradually tend towards the target steady-state distribution, unaffected by environmental drift, device errors, and noise. This reduces external offline trial and error, realizing a functional extension from "optimization only" to "controllable sampling / generation," achieving a balance between stable reproduction and adjustable exploration-convergence.

[0026] Details of one or more embodiments of this application are set forth in the following drawings and description to make other features, objects and advantages of this application more readily apparent. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other embodiments can be obtained based on these drawings without creative effort.

[0028] Figure 1 This is a flowchart illustrating the training method for the coherent Ising machine provided in the embodiments of this application.

[0029] Figure 2 This is a schematic diagram of the structure of the computing device based on the coherent Ising machine provided in this embodiment.

[0030] Figure 3 This is a schematic diagram illustrating a specific application principle of the computing device based on the coherent Ising machine provided in this embodiment.

[0031] In the picture: 10. Coherent Ising Machine; 11. Laser Pulse Generator; 12. Control Module; 121. Pump Modulation Unit; 122. Saturation Control Unit; 13. Feedback Module; 131. Measurement Unit; 132. Calculation Unit; 133. Feedback Injection Unit; 134. Sampling Unit; 20. Noise Controllable Injector; 30. Training Controller. Detailed Implementation

[0032] Embodiments of this application will now be described in more detail with reference to the accompanying drawings. While some embodiments of this application are shown in the drawings, it should be understood that embodiments of this application can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the embodiments of this application. It should be understood that the accompanying drawings and embodiments are for illustrative purposes only and are not intended to limit the scope of protection of this application.

[0033] The Coherent Ising Machine 10 (CIM) was initially proposed for solving combinatorial optimization problems, especially NP-hard problems of the Ising / QUBO form. Its core hardware consists of a set of decoherent optical parametric oscillators (DOPO) pulses, coupled through a measurement-feedback (MFB) mechanism.

[0034] Currently, traditional CIM (Continuous Imaging Model) primarily aims to find the optimal solution / low-energy state. Hardware parameters (pump, saturation, gain, coupling) are mostly fixed or empirically tuned, causing the steady-state sample distribution of optical pulses to vary with environmental drift and device errors. This makes it difficult to form a controllable sampler with a "target Gibbs distribution." The Gibbs distribution is a statistical distribution describing the probability of a system occurring in a given state, related to the state energy.

[0035] Secondly, the feedback link of the Coherent Ising Machine 10 lacks a closed-loop self-configuration mechanism between itself and the physical parameters: existing technologies often treat the coupling matrix or bias field as external settings, lacking a hardware closed loop that automatically updates physical parameters such as pump / coupling / external field / gain based on measurement statistics (amplitude moment, correlation moment) within the device. This results in high costs of repeated deployments and poor adaptability. Furthermore, noise and temperature (diffusion intensity) lack controllable injection and equivalent modeling: real-world systems contain quantum noise, probe noise, and electronic noise. Without a controllable noise injection / equivalent temperature adjustment module, it is often difficult to achieve stable reproduction and adjustable exploration-convergence balance across different tasks. The bias field includes the external field or bias, which can be implemented by external drive and feedback bias, respectively.

[0036] To address the aforementioned issues, firstly, the first embodiment of this application provides a training method for a coherent Ising machine 10. The coherent Ising machine 10 includes a laser pulse generator 11, which generates N optical pulses with continuous amplitudes, where N is a natural number, and the N continuous amplitudes constitute an N-dimensional continuous amplitude vector. The continuous amplitudes can be in-phase amplitudes or quadrature components of the optical pulses.

[0037] like Figure 1 As shown, the training method provided in this application includes the following steps: Step S11: Inject noise into the coherent Ising machine 10 so that, under the combined action of the physical parameters of the coherent Ising machine 10 and the noise, the continuous amplitude vector evolves to a steady-state distribution that follows the Gibbs distribution according to the Langevin dynamics evolution rule. The physical parameters include pump parameters, saturation parameters, coupling coefficients, bias field and gain.

[0038] The Langevin dynamics evolution rule describes the Brownian motion evolution of particles as depicted in Langevin dynamics. In Langevin dynamics, particle motion is influenced not only by deterministic forces as described in classical mechanics but also by a random force (noise term). Based on this, under the combined influence of physical parameters and noise, the change (or amplitude) of the continuous amplitude of the light pulse in this application is equivalent to the Brownian motion of the particle. The pump parameters, saturation parameters, coupling coefficient, bias field, and gain are equivalent to deterministic forces applied to the light pulse, while the noise is the random force applied to the light pulse. This allows the coherent Ising machine 10 to be modeled as an optical Langevin machine, whose output light pulse state is always a continuous optical amplitude variable. Therefore, the continuous amplitude of the N light pulses output by the coherent Ising machine 10 will evolve to a steady-state distribution following a Gibbs distribution, thus allowing the coherent Ising machine 10 to be considered a physically realized energy model sampler. Therefore, the coherent Ising machine 10 provided in this application can be trained like an energy model, and the energy parameters for training such a coherent Ising machine 10 can be completely physicalized as its physical parameters, which provides a training basis for this application to train the coherent Ising machine 10.

[0039] Step S12: After using the continuous amplitude vector collected under the current steady-state distribution as negative phase amplitude data, one or more physical parameters are updated based on the update amount obtained from the deviation between the negative phase amplitude data and the positive phase amplitude data, so that the continuous amplitude vector reaches the target steady-state distribution. The positive phase amplitude data comes from a pre-set training dataset. Under the target steady-state distribution, the continuous amplitude vector of the optical pulse output by the coherent Ising machine 10 will evolve to a state closer to the positive phase amplitude data. During the training process, dynamic updates of physical parameters are realized, and the actual dynamics and steady-state distribution of the coherent Ising machine 10 (or optical Langevin machine) are controllably adjusted during the updating of physical parameters. Dynamic updates of physical parameters also avoid repeated deployment of the coherent Ising machine 10 hardware, reduce the deployment cost of the coherent Ising machine 10, and improve its adaptability. Where the gain is a hardware-calibrated and fixed value, there is no need to update the gain.

[0040] In some implementations, when the positive phase amplitude data comes from the training dataset, it can be based on real training samples extracted from the Coherent Ising Machine 10 training dataset (such as a mini-batch consisting of N selected observation samples). The quantized data, which reflects the gradient characteristics of the real training sample, is obtained by calculating the positive phase gradient of the real training sample and is used to compare with the negative phase amplitude data to update the physical parameters of the coherent Ising machine.

[0041] Specifically, the positive phase amplitude data can be selected from the training dataset in the following way: The training objective of the Coherent Ising Machine 10 can be to minimize the divergence KL between the distribution of the real sample data in the training dataset and the evolution model of the Coherent Ising Machine 10: KL( When ), the training objective for the coherent Ising machine 10 is equivalent to maximizing the log-likelihood. The log-likelihood in the sample... The gradient on the data can be decomposed into two parts, where the term related to the data distribution is called the positive phase: Select one from the training dataset mini-batch consisting of observation samples The empirical estimate of positive phase amplitude data is defined as:

[0042] The calculation of this quantity does not require sampling from the model distribution; the steps are as follows: 1) Take a mini-batch from the training dataset .

[0043] 2) Perform forward model calculations for each observation sample to obtain the corresponding energy scalar. .

[0044] 3) For energy scalars For model parameters By taking the derivative of (physical parameters and noise), we can obtain the gradient. .

[0045] 4) Gradient of all observation samples within the mini-batch Take the average value, then add a negative sign to the average value to obtain the positive phase gradient estimate. Based on this positive gradient estimation Select the positive phase amplitude data.

[0046] From an optimization perspective, positive phase amplitude data helps the model reduce the energy at the observed sample points. This increases the probability density of observed sample data in the model. The calculation of positive phase amplitude data can be completely determined by the microstructure of the observed samples and energy function, ultimately achieving the training objective of minimizing KL divergence and maximizing log-likelihood.

[0047] The negative phase amplitude data in this application comes from the sampled samples under the natural evolution steady-state distribution of the coherent Ising machine 10 under the current physical parameters, without the need for computer simulation.

[0048] As can be seen, the training method for the coherent Ising machine 10 provided in this application, after injecting noise into the coherent Ising machine 10, will determine the continuous amplitude of the light pulses output by the coherent Ising machine 10 by both its physical parameters and the noise. The change (or amplitude) of the continuous amplitude of the light pulses will be equivalent to the Brownian motion of tiny particles in Langevin dynamics. The pump parameters, saturation parameters, coupling coefficient, bias field, and gain are equivalent to deterministic forces applied to the light pulses, while the noise is a random force applied to the light pulses. This allows the coherent Ising machine 10 to be modeled as an optical Langevin machine, whose output light pulse state is always a continuous optical amplitude variable. Therefore, the continuous amplitude of the N light pulses output by the coherent Ising machine 10 will evolve to a steady-state distribution following a Gibbs distribution, thus allowing the coherent Ising machine 10 to be considered a physically realized energy model sampler.

[0049] Therefore, the coherent Ising machine 10 obtained in this application can be trained like an energy model, and the energy parameters of such a coherent Ising machine 10 can be fully physicalized into its own physical parameters. Based on this, after using the continuous amplitude vector collected under the current steady-state distribution as negative phase amplitude data, this application can update the physical parameters based on the update amount obtained from the deviation between the negative and positive phase amplitude data, so that the steady-state distribution reaches the target steady-state distribution. Thus, during the operation of the coherent Ising machine 10, the physical parameters can be updated in real time according to the collected continuous amplitude, so that the amplitude of the light pulse gradually tends towards the target steady-state distribution, unaffected by environmental drift, device errors, and noise, reducing external offline trial and error, realizing the functional extension from "optimization only" to "controllable sampling / generation," and achieving a balance between stable reproduction and adjustable exploration-convergence.

[0050] In this application, the step of obtaining the update amount based on the deviation between the negative phase amplitude data and the positive phase amplitude data includes: After calculating the first statistic for the negative phase amplitude data and the second statistic for the positive phase amplitude data, an update quantity is generated based on the difference between the first and second statistics. Both the first and second statistics include one or more of the following statistics: The first moment, which is a statistic required to update the bias field, represents the amplitude of the i-th optical pulse. The mean of i∈N. In this case, the first moment can be used... Indicates. If If it belongs to the first statistic, then the amplitude For continuous amplitude values ​​in negative phase amplitude data, if If it belongs to the second statistic, then the amplitude This refers to the initial continuous amplitude or the target continuous amplitude in the positive phase amplitude data.

[0051] The second moment, a statistic required to update the pump parameters, represents the amplitude of the i-th optical pulse. The average of the squares. The second moment can be expressed as... Similarly, the amplitude varies when the second moment belongs to different statistics. This corresponds to the amplitude in different amplitude data.

[0052] The second-order correlation moment is a statistic required to update the coupling coefficients, representing the amplitude of the i-th optical pulse. and the amplitude of the j-th light pulse The average of the products between them, j∈N. The second-order correlation moment can be expressed as: .

[0053] The fourth moment is a statistic required to update the saturation parameters, representing the average of the fourth power of the amplitude of the i-th light pulse. The fourth moment can be expressed as: .

[0054] The amplitude energy correlation term is a statistic required for updating the gain. It is the average energy of the coupling and bias terms. The coupling term is the total energy of the interaction between all pulse pairs formed by the i-th and j-th optical pulses. The bias term is the energy of the i-th optical pulse under the influence of the bias field. The amplitude energy correlation term satisfies the following relationship: .

[0055] Among them For coupling terms, The coupling coefficient between the i-th and j-th optical pulses represents the weight of the interaction between the pulse pairs. This is a bias term.

[0056] In some implementations, the above-mentioned statistical mean can be achieved by a sliding window average, a segmented accumulator, or a parallel accumulator.

[0057] This application updates and adjusts the corresponding physical parameters based on the difference between the first and second statistics. Essentially, it treats the difference as an approximate gradient, driving the physical parameters to adjust in a direction that reduces the evolution error of the coherent Ising machine 10 system, thereby achieving a stable distribution of the optical pulses. Since the difference simultaneously includes the system response in both positive and negative directions, it can automatically suppress noise and cancel disturbances, making the updated values ​​more stable. Therefore, the essence of this update process is to change the physical parameters of the optical system, thereby altering the dynamics and steady-state distribution of the coherent Ising machine 10 system, achieving effects such as controllable amplitude distribution, improved stability, and reduced energy consumption / delay.

[0058] In this application, the update quantity includes the update quantity used to update the bias field quantity. The steps for generating the update value based on the difference between the first and second statistics include: Step S21: Obtain the first moment of the first statistic The first moment of the second statistic The noise intensity D and the learning rate required to update the bias field. The learning rate is a parameter that controls the update step size. In this application, the required learning rate can be configured in a hardware register set by the coherent Ising machine 10.

[0059] Step S22: After calculating and After calculating the difference between them, this difference will be combined with the gain and learning rate. The product of the three factors, including the reciprocal of D, is added to the bias field before the update to obtain the update value. .

[0060] Set the bias field before the i-th optical pulse update to be Based on the above steps S21 to S22, the updated quantity of this application The update relation is satisfied as follows: = + ( - ).

[0061] Then, based on the update amount obtained above... This allows for the updating of the bias field, thereby driving the steady-state distribution of the coherent Ising 10 system toward the target steady-state distribution.

[0062] In this application, the gain in the update amount of gain g can be the updated gain, the gain fixed by the system, or the gain before the update. The specific setting depends on the actual needs, and this application does not impose a unique limitation.

[0063] The update quantities in this application also include update quantities used to update pump parameters. Then, the step of generating the update value based on the difference between the first statistic and the second statistic further includes: Step S31: Obtain the second moment of the first statistic The second moment of the second statistic The noise intensity D and the learning rate required to update the pump parameters. .

[0064] Step S32: After calculating After calculating the difference between them, the obtained difference is then compared with the learning rate. The product of the inverse of 2D and the pump parameters before the update is added to obtain the update amount. .

[0065] If the pump modulation unit 121 of the coherent Ising machine 10 modulates the pump parameters separately for each optical pulse, the pump parameters before the update are set to... If the pump modulation unit 121 modulates a global pump parameter for all optical pulses, then the pump parameter before the update is set to p. Based on the above steps S31 to S32, the update amount in this application... The update relation is satisfied as follows: = + ( - ).

[0066] If the global pump parameter p (all) is used ≡ p), then the update amount The update relation is satisfied as follows: = + ( - ).

[0067] This application is based on the update quantity obtained above. This enables the updating of pump parameters, thereby driving the steady-state distribution of the coherent Ising 10 system toward the target steady-state distribution.

[0068] The update amounts in this application also include update amounts used to update the coupling coefficients. Then, the step of generating the update value based on the difference between the first statistic and the second statistic further includes: Step S41: Obtain the second-order correlation moments of the first statistic The second correlation moments of the second statistic The noise intensity D and the learning rate required to update the coupling coefficients. .

[0069] Step S42: After calculating and After calculating the difference between them, the obtained difference is then compared with the learning rate. The product of the inverse of 2D and the gain is added to the coupling coefficient before the update to obtain the update amount. .

[0070] The coupling coefficient before the update is set to The coupling coefficient is the coefficient required by the feedback module 13 of the coherent Ising machine 10 to calculate the feedback amount for feedback adjustment of the optical pulse. Based on the above steps S41 to S42, the update amount... The update relation is satisfied as follows: = + ( - ).

[0071] Therefore, this application is based on the update quantity obtained above. This allows for the updating of coupling coefficients, thereby driving the steady-state distribution of the coherent Ising machine 10 system toward the target steady-state distribution.

[0072] In some implementations, if the hardware / modeling requirements for the coherent Ising machine 10 are set, the coupling coefficients of all pulse pairs are... The coupling matrix J formed is symmetric and When =0, based on the update amount After updating the coupling coefficients, this application can also perform symmetric projection and zero diagonal constraint operations on the coupling coefficients. The symmetric projection operation is the operation of projecting the coupling matrix J onto the symmetric matrix space, and the zero diagonal constraint operation is the operation of setting the elements on the diagonal of the coupling matrix J to zero.

[0073] The symmetric projection operation performed after each update is as follows: J← , It is the transpose of the coupling matrix.

[0074] And after each update, the zero diagonal constraint is executed as follows: diag(J)←0, which sets all diagonal elements to zero, where diag(J) is the main diagonal element taken from the coupling matrix J, such as: , …, Therefore, this application ensures that the coupling matrix J always falls within the feasible region by performing symmetric projection operations and zero diagonal constraint operations.

[0075] The update quantity in this application also includes the update quantity used to update the saturation parameter. Then, the step of generating the update value based on the difference between the first statistic and the second statistic further includes: Step S51: Obtain the fourth moment of the first statistic The fourth moment of the second statistic The noise intensity D and the learning rate required to update the saturation parameters. .

[0076] Step S52: After calculating and After calculating the difference between them, the obtained difference is then compared with the learning rate. The product of the inverse of 4D and the saturation parameter before the update is added to obtain the update amount. .

[0077] The saturation parameter before the i-th light pulse update is set to Based on the above steps S51 and S52, the update quantity in this application The update relation is satisfied as follows: = + ( - ).

[0078] Therefore, this application is based on the update quantity obtained above. Achieve saturation parameters for each optical pulse The update drives the steady-state distribution of the coherent Ising machine 10 system to tend toward the target steady-state distribution.

[0079] The update amount in this application also includes the update amount used to update the gain. Then, the step of generating the update value based on the difference between the first statistic and the second statistic further includes: Step S61: Obtain the amplitude energy correlation term of the first statistic The amplitude energy correlation term of the second statistic The noise intensity D and the learning rate required to update the gain. ,in, Let be the coupling coefficient between the i-th optical pulse and the j-th optical pulse. For coupling terms, This is a bias term.

[0080] Step S62: After calculating and After calculating the difference between them, the obtained difference is compared with the reciprocal of D and the learning rate. The product of the two is summed with the gain before the update to obtain the update amount. .

[0081] Let the gain before the update be g. Based on steps S61 and S62 above, the update amount is... The update relation is satisfied as follows: =g+ [ - ].

[0082] Therefore, this application is based on the update quantity obtained above. This allows for the updating of the gain of each optical pulse, thereby driving the steady-state distribution of the coherent Ising 10 system towards the target steady-state distribution.

[0083] In some implementations, in most application scenarios, the gain g can be calibrated and fixed by hardware, which makes the training pump parameters, coupling coefficients, saturation parameters, and external bias field parameters more stable.

[0084] In the training method provided in this application, when it is necessary to train and update the gain through the above steps S61 to S62, the method provided in this application may further include the following steps: Through update volume During the gain update process, the magnitude of gain changes is constrained based on predefined rules. These rules include magnitude constraints or regularization terms. Magnitude constraints limit the range of gain values, while regularization terms are used to adjust the calculation of the update amount. A penalty term is added during the process to suppress excessive changes in gain. This application improves the stability of the gain value by setting amplitude constraints or regularization terms, thereby improving the stability of the entire training process.

[0085] The statistical quantities and update relationships of the above-mentioned items in this application are derived based on the following method and process: 1. The realization from the coherent Ising machine 10 to the optical Langevin machine From the perspective of the dynamics of the continuous amplitude realization of optical pulses by the Coherent Ising Machine 10 (hereinafter referred to as CIM) in this application, the system state of the Coherent Ising Machine 10 is not a discrete spin, but a continuous amplitude variable for each optical pulse. These amplitudes evolve under the combined effects of nonlinear gain, loss, coupling, and noise, naturally forming a stochastic dynamical system. That is to say, the noisy Coherent Ising Machine 10 is equivalent to a Langevin system in the continuous amplitude space, and its steady-state distribution is a Gibbs distribution, so it can be regarded as an optical Langevin machine (OLM). From this perspective, CIM is not only an optimizer, but also a physical sampler. CIM can be trained like an energy model (EBM), and the training parameters can be fully physicalized.

[0086] 2. The Continuous Amplitude Dynamic Evolution Principle and Physical Parameters of CIM 2.1 CIM System State and Variable Definitions: Consider a CIM system consisting of N light pulses, and denote its N-dimensional continuous amplitude vector (using...). (This indicates that) satisfies the following relation: = ∈ .

[0087] Where R represents Each component is a real number.

[0088] In this application, It is always treated as a continuous random variable and no discretization is performed (e.g., sign( (Do not enter the dynamics and training objective). i is between 1 and N.

[0089] 2.2 Unified notation of CIM physical parameters: The trainable or adjustable physical parameters are denoted as θ: θ = {p, k, J, h, g}.

[0090] in: p= Pump parameters (net linear gain control parameters relative to the threshold, or global p can be used).

[0091] k= : Saturation parameter (nonlinear term of laser pulse generator 11, typically >0).

[0092] J∈ The coupling matrix implemented by the feedback link (usually symmetric and constrainable) =0).

[0093] h∈ : For external field or bias injection (can be achieved by feedback bias or external drive).

[0094] g∈ Feedback gain (overall coupling strength or amplification factor, often calibrated by hardware or updated through training).

[0095] 2.3 Physicalized Effective Potential Energy Function To match the amplitude drift term of CIM, this application adopts the following fully physicalized potential energy function: (1) And assume J is symmetric [if it is not symmetric, then use its symmetric part]. All of the above physical parameters can be configured by the corresponding CIM hardware.

[0096] By controlling the noise injection intensity, the CIM system can achieve an effective diffusion intensity. Thus, the Gibbs distribution is obtained in steady state.

[0097] 3. Optical Langevin Dynamics and Steady-State Distribution 3.1 Langevin stochastic differential equations (continuous amplitude, not discretized) After considering the equivalent effects of various noise types such as quantum noise, detector noise, and electronic noise, the amplitude dynamics of CIM can be modeled as an overdamped Langevin equation: da(t) = (a(t))dt+ dW(t), (2) Where D>0 represents the diffusion intensity (effective temperature), W(t) represents standard Brownian motion, and t is the time parameter. This equation describes the variation of the amplitude of the entire CIM system over time. The evolution of the amplitude vector is driven by the negative gradient force (the term before the plus sign is the gradient term, and the term after is the noise term), evolving towards the direction of lower potential energy (higher probability), while also being affected by random noise. The disturbances bring exploration and diversity. Overdamped systems mean that the evolution of the system is influenced by damping forces, rather than inertia.

[0098] Taking the gradient of (1) yields the component-wise drift form: d =[( ) ]dt+ (t), (3) That is, the continuous amplitude of the i-th pulse (in-phase component or equivalent real amplitude).

[0099] 3.2 Fokker–Planck Equation (This application is a partial differential equation describing the evolution of the probability density function over time during the random evolution of amplitude) The corresponding probability density pθ(a, t) satisfies the following relationship: (4) 3.3 Steady-state Gibbs distribution Under the condition of zero probability flow (detailed balance), the CIM system exhibits a steady-state distribution: (5) Therefore, CIM naturally implements Gibbs sampling in a continuous amplitude space. Any discrete readout (such as sign( ()) Do not enter into dynamics and training derivation. This is the normalization constant.

[0100] 4. Training objective: Maximum likelihood and positive / negative phase decomposition. Suppose the training data comes from the target steady-state distribution amplitude samples. ( (i.e., positive phase amplitude data), the training objective is to minimize the divergence KL between positive phase amplitude data and negative phase amplitude data samples: KL( ).

[0101] The training objective is equivalent to maximizing the log-likelihood: (6) From formula (5), we can obtain: .

[0102] Taking the gradient with respect to θ yields the gradient decomposition of the classical energy model: + (7) Among them, the positive phase amplitude data is Positive phase amplitude data can reduce the energy at the true target amplitude data, making the continuous amplitude corresponding to the data have a higher probability. Negative phase amplitude data allows the CIM system model to assign higher energy or lower probability to "unrealistic" or "unreasonable" amplitude sample data. Negative phase amplitude data is expected to be obtained through Langevin dynamics sampling (physical sampling or numerical sampling) using an optical Langevin machine (OLM).

[0103] 5. Physics Parameter Training: Derivation from Potential Energy to Explicit Update Formula The θ in formula (7) is specified as follows: θ = {p, k, J, h, g}.

[0104] 5.1 Positive / Negative Phase Samples and Labels Let the training samples of the positive phase amplitude data in a training subset mini-batch be... , The data sample number is used; the training samples for negative phase amplitude data are... (Evolved from CIM), where m is the sampling number. The expected value is approximated by the sample mean: := , := .

[0105] And it is updated using a gradient ascent method: θ ←θ+ , This represents the learning rate corresponding to the physical parameters.

[0106] 5.2 Updating the pump parameter p From formula (1), we get: .

[0107] Substituting into formula (7), we get: ← = + ( - (8).

[0108] If a global pump p is used (all pi ≡ p), then: = + ( - (9).

[0109] 5.3 Update of saturation parameter k (and maintain) >0): .

[0110] From formula (7), we get: ← = + ( - (10).

[0111] Feasibility (Projection / Parameterization Recommendations): To ensure the potential energy is normalizable and to avoid amplitude divergence, it is typically necessary to... >0 (If it is less than zero, the probability distribution cannot be normalized, and the model is unstable). Projection can be used: ← max( , ε), that is, after each update, if If it is less than a positive number ε, then a forced setting is required. For ε. Or use logarithmic parameterization instead. Performing unconstrained optimization means using a new unconstrained variable. The saturation parameters are parameterized. The stability of the CIM steady-state distribution is improved through these methods.

[0112] 5.4 Coupling Matrix Update (MFB weights) From formula (1): .

[0113] From formula (7), we can obtain: ← = + ( - ,(11)。

[0114] If the coherent Ising machine 10 hardware / modeling requires the coupling coefficient of all pulse pairs... The coupling matrix J formed is symmetric and When the value is 0, this application can also perform symmetric projection and zero diagonal constraint operations. This can be done after each update: J← ,diag(J)←0,(12).

[0115] Set all diagonal elements to zero, where diag(J) is the main diagonal element taken from the coupling matrix J, such as: , …, Therefore, this application ensures that the coupling matrix J always falls within the feasible region by performing a projection operation and a zero diagonal constraint.

[0116] 5.5 Update of the field / bias h From formula (1): .

[0117] Similarly, the calculation yields: ← = + ( - (13).

[0118] 5.6 Feedback Gain Update (Optional) From formula (1): .

[0119] Similarly, the update formula for the feedback gain can be obtained: g← =g+ [ - ],(14)。

[0120] Therefore, the above derivation process provides a theoretical basis and implementation method for training and adjusting the physical parameters of the coherent Ising machine 10 in this application.

[0121] In this application, the step of updating physical parameters based on the update amount includes: Update amount used for pump parameters The data is sent to the pump modulation unit 121 of the coherent Ising machine 10 so that the pump modulation unit 121 can update the data based on the amount of data. The optical pulse is modulated. The pump modulation unit 121 can be based on the update amount. This allows for programmable modulation of the pump parameters (such as pump intensity and pump power) for each optical pulse, enabling the modulation of different optical pulses. Or, the modulation update of global p.

[0122] When updating physical parameters based on the update amount, the update amount used to update saturation parameters can also be included. The data is sent to the saturation control unit 122 of the coherent Ising machine 10 so that the saturation control unit 122 can update the data based on the amount of data. The saturation parameter is controlled within a stable range. The saturation control unit 122 can maintain the saturation parameter within a stable range through intracavity nonlinearity, controllable losses, or equivalent electronic compensation.

[0123] When updating physical parameters based on the update amount, the update amount used to update the gain can also be used. Update amount of the coupling coefficient And the update amount of the updated bias field. The data is sent to the feedback module 13 of the coherent Ising machine 10, so that the feedback module 13 updates the continuous amplitude and amount of the optical pulse according to the current steady-state distribution. Update volume and update volume The obtained feedback quantity is used to adjust the optical pulse. The feedback quantity includes the total amount of coupling influence of each optical pulse on other optical pulses (e.g., the total coupling influence of the i-th optical pulse is...). and the bias field quantity it receives (Outside / Offset).

[0124] Therefore, this application achieves fully physical training by distributing the corresponding update quantity to the corresponding hardware of the coherent Ising machine 10 and updating and adjusting the corresponding physical parameters within the hardware, thereby reducing external offline trial and error and resisting hardware drift.

[0125] In addition, the method provided in this application also includes the following steps: Negative-phase amplitude data and positive-phase amplitude data are input to the calculation unit 132 of the feedback module 13. The calculation unit 132 calculates the first and second statistics, and generates an update value based on the difference between the first and second statistics. Thus, by using the calculation unit 132 (such as a dedicated MAC or cross-array) configured within the coherent Ising machine 10 itself for statistical calculations along a fixed data path, the involvement of a general-purpose processor is reduced, lowering data handling and storage overhead. This achieves lower latency and better energy efficiency at the same sampling throughput.

[0126] A second embodiment of this application provides a computing device based on a coherent Ising machine, which is used to execute the training method of the coherent Ising machine provided in the first embodiment of this application. The computing device includes a coherent Ising machine 10, a noise-controlled injector 20, and a training controller 30.

[0127] The coherent Ising machine 10 includes a laser pulse generator 11, a control module 12, and a feedback module 13, with the control module 12 and feedback module 13 connected to the laser pulse generator 11. The laser pulse generator generates N optical pulses with continuous amplitudes. The control module 12 controls the pump and saturation parameters of the optical pulses, and the feedback module 13 adjusts the optical pulses based on feedback values ​​obtained from the continuous amplitude, gain, coupling coefficient, and bias field of the optical pulses.

[0128] The noise controllable injector 20 is connected to the laser pulse generator 11 and the feedback module 13 respectively. The noise controllable injector 20 is used to generate controllable noise and inject the noise into the laser pulse generator 11 or the feedback module 13, that is, to inject into the corresponding optical path to form a disturbance, thereby realizing the setting and adjustment of the diffusion intensity D.

[0129] The training controller 30 is connected to the control module 12 and the feedback module 13. The training controller 30 is used to obtain the update amount based on the deviation between the collected negative phase amplitude data and the positive phase amplitude data, and sends the update amount to the control module 12 and the feedback module 13. After receiving the update amount for updating the corresponding physical parameters, the control module 12 and the feedback module 13 update the corresponding physical parameters.

[0130] Therefore, this application adds a closed loop of "training controller 30 + noise controllable injector 20 + coherent Ising machine 10 and related modulation device" inside the computing device, so that the computing device can automatically update and send out physical parameters according to the statistical quantity of the measured negative phase amplitude data sample, thereby making the steady-state distribution of the coherent Ising machine 10 closer to the target distribution or the target sample statistics.

[0131] In some embodiments, the noise controllable injector 20, training controller 30, etc., can be directly configured inside or outside the coherent Ising machine 10, and this application does not impose a unique limitation on this.

[0132] The laser pulse generator 11 may include an optical parametric oscillator array or a time-division multiplexed DOPO pulse ring cavity. The time-division multiplexed DOPO pulse ring cavity may include a ring cavity, a nonlinear crystal (e.g., PPLN), a coupler, optical isolation and frequency stabilization components, etc. Figure 3 (As shown). PPLN stands for Periodically Poled Lithium Niobate. In some implementations, the state of the current optical pulse can be measured by a "sensor" such as a second harmonic pulse. After calculation, the optical parametric oscillator (DOPO) is controlled by a pump modulation unit to determine how the next optical pulse is generated. This is the key feedback loop of the coherent Ising machine.

[0133] The noise controllable injector 20 in this application may include a noise source (thermal noise / band-limited white noise / pseudo-random Gaussian), a programmable filter module, and an amplitude calibration module. The programmable filter module is connected between the noise source and the amplitude calibration module, and the amplitude calibration module is connected to the laser pulse generator 11 and the feedback module 13. The programmable filter module adjusts the noise frequency characteristics, and the amplitude calibration module can adjust the noise amplitude or power. The amplitude calibration module injects the adjusted noise into the feedback module 13, the laser pulse generator 11, and other hardware according to the pulse time slots, making the equivalent diffusion intensity D adjustable.

[0134] The above update volume includes update volume Update volume Update volume Update volume and update volume The control module 12 includes a pump modulation unit 121 and a saturation control unit 122. The pump modulation unit 121 modulates the optical pulses. The pump modulation unit 121 includes a pump laser and an electro-optic modulator or acousto-optic modulator for modulating the pump intensity per node (or per pulse time slot). The saturation control unit 122 controls the saturation parameters within a stable range. In some embodiments, the saturation control unit 122 may include a controllable cavity loss device (VOA), a controllable phase / polarization control and compensation circuit, or an equivalent electrically controlled saturation compensation, so that... It is within a stable range and can be calibrated under manufacturing deviations.

[0135] Both the saturation control unit 122 and the pump modulation unit 121 are connected to the training controller 30. The training controller 30 updates the quantity... The update amount is sent to the pump modulation unit 121. The updated amount is sent to the saturation control unit 122. The update amount and update volume Send to feedback module 13.

[0136] Therefore, this application achieves fully physical training by distributing the corresponding update quantity to the corresponding hardware of the coherent Ising machine 10 and updating and adjusting the corresponding physical parameters within the hardware, thereby reducing external offline trial and error and resisting hardware drift.

[0137] The coherent Ising laser 10 also includes a measurement unit 131, which measures the continuous amplitude of the optical pulses. The feedback module 13 includes a calculation unit 132 and a feedback injection unit 133. The measurement unit 131 is connected to the laser pulse generator 11 and measures the continuous amplitude of the optical pulses. Specifically, the measurement unit 131 can perform in-phase or heterodyne measurements on the continuous amplitude of each optical pulse and output an analog electrical signal.

[0138] The calculation unit 132 is connected to the measurement unit 131 and the training controller 30 respectively. After receiving the continuous amplitude measured by the measurement unit 131, the calculation unit 132 generates a feedback quantity. + It also supports online updates and constrained projections (symmetry, zero diagonal, amplitude clipping, etc.) of coupling coefficients and bias field quantities.

[0139] like Figure 2 and Figure 3 As shown, the feedback injection unit 133 is connected to the calculation unit 132 and the laser pulse generator 11. The feedback injection unit 133 is used to inject feedback quantity into the laser pulse generator 11 to perform feedback adjustment of the optical pulse. The feedback injection unit 133 may include a digital-to-analog converter (DAC) and a pulse modulator (such as an electro-optic modulator, amplitude-phase modulator, or direct optical injector) connected to each other. The DAC is used to convert the feedback quantity in digital signal form into an analog signal. The pulse modulator uses this analog signal (such as a voltage signal) to change the phase or amplitude of the corresponding optical pulse in real time. The modulated optical pulse participates in the next round of training process to complete the closed-loop iteration. The feedback injection unit 133 can convert the feedback quantity containing noise into an analog voltage and apply it to the optical pulse. When a gain g is set, the pulse modulator will... Injecting optical pulses enables coupling and external field / bias control.

[0140] The training controller 30 can calculate the first and second statistics through the computing unit 132, and generate an update value based on the difference between the first and second statistics. This reduces the involvement of a general-purpose processor and lowers data transfer and storage overhead. It achieves lower latency and better energy efficiency at the same sampling throughput. The training controller 30 can calculate the statistics and update value within the computing unit 132 using a fixed data path, and drive various physical modulation devices (pump modulation, DAC scaling, coupling weight register), thereby changing the actual dynamics and steady-state distribution of the CIM optical system.

[0141] The feedback module 13 also includes a sampling unit 134, which is connected between the measurement unit 131 and the calculation unit 132. The sampling unit 134 is used to collect amplitude data from the continuous amplitude measured by the measurement unit 131 and send it to the calculation unit 132 during the evolution of the coherent Ising machine 10 under the current physical parameters and noise.

[0142] The sampling unit 134 can also send the amplitude data obtained by aligning the measured continuous amplitude with the original N optical pulses in time to the calculation unit 132, so as to ensure that when the calculation unit 132 performs subsequent calculation of feedback and injection feedback, it can correspond one-to-one with the corresponding optical pulse, and it is not easy for data identification to be misaligned, thereby improving the accuracy of feedback.

[0143] In some implementations, such as Figure 2 As shown, the measurement unit 131 can be configured externally to the feedback module 13 and connected to the laser pulse generator 11. In other embodiments, the measurement unit 131 can also be configured within the laser pulse generator 11, such as... Figure 3 As shown ( Figure 3 (This is a schematic diagram illustrating the continuous amplitude evolution principle when the DOPO pulse ring cavity is connected to various components). Measurement unit 131 can be configured in the time-division multiplexed DOPO pulse ring cavity, and measurement unit 131 is connected between the nonlinear crystal of the DOPO pulse ring cavity and the sampling unit 134. The specific configuration of measurement unit 131 is not limited in this application.

[0144] In some embodiments, when the continuous amplitude measured by the measurement unit 131 is an analog electrical signal, the sampling unit 134 may include an analog-to-digital converter (ADC) and a clock synchronization and alignment circuit. The output of the ADC is connected to the input of the clock synchronization and alignment circuit, which is connected to the calculation unit 132. The ADC is used to digitize the analog electrical signal to obtain a digital amplitude signal, and the clock synchronization and alignment circuit is used to send the amplitude data obtained by aligning the digital amplitude signal with the original N optical pulses in time to the calculation unit 132.

[0145] The computing unit 132 includes a control chip and computing devices. The control chip is connected to both the measurement unit 131 and the training controller 30. The control chip includes either a Field-Programmable Gate Array (FPGA) or an Application-Specific Integrated Circuit (ASIC) for controlling the computation process. The computing devices are connected to the control chip and include either a Multiply-Accumulate MAC array or a Cross Array for performing multiplication and addition operations on feedback quantities, first statistics, second statistics, and update quantities. The gain g is a coefficient multiplied on the feedback quantity before the digital-to-analog converter, used for online adjustment of the feedback strength. Real-time adjustment of g via the FPGA can maintain stability and accelerate convergence at different operating points.

[0146] In some implementations, the operation flow of the computing device provided in this application (embodied in the control flow of the computing device, implemented by state machines and timing control in each hardware component) may include the following steps: Step 1) Initialization and Calibration: Read initial parameters such as p, k, J, h, g, and D from the parameter storage. Temperature control / frequency stabilization enters a stable state. Inject noise of a certain intensity (e.g., voltage noise 160mV), set the target spectral density of the noise, and calibrate the equivalent diffusion coefficient D of the equipment based on the target spectral density of the noise and the system response.

[0147] Step 2) Sampling Operation: The DOPO pulse train evolves in the optical ring cavity. The continuous amplitude signal of the output is measured. Sampling unit 134 performs ADC acquisition and clock alignment settings.

[0148] Step 3) Negative phase statistics: Training the controller to perform 30 tests within the steady-state sampling window. calculate , Equal statistics.

[0149] Step 4) Positive Phase Statistics: A sample window of positive phase amplitude data can be generated from target data samples mapped or cached from external input, and the corresponding statistics can be calculated (or the statistical target can be directly provided by external input): , .

[0150] Step 5) Calculation of parameter update amount: Generate according to the update formula above. , , , , .

[0151] Step 6) Parameter delivery and modulation: The updated parameters are written into the pump modulation, coupling weights, feedback injection, and noise injection to change the dynamics of the real physical system.

[0152] Step 7) Iteration: Repeat steps 2) to 6) until a preset target steady-state distribution is reached. For example, the statistical error is less than a threshold, meaning the difference between the statistic of the current sampled amplitude data sample and the statistic of the positive phase amplitude data sample is less than a preset threshold. The smaller the difference, the closer the statistical characteristics of the sampling results are to the target steady-state distribution. Another example is reaching a steady-state distribution distance index decrease, which measures the "distance" or "difference" between the current sampled distribution and the target distribution. The index is the KL divergence mentioned in the first embodiment of this application.

[0153] The computing device provided in this application enables the steady-state distribution of the coherent Ising machine 10 to be controllable and calibrable: by introducing a noise controllable injector 20 and a self-configured closed-loop update of physical parameters, the computing device can form a controllable Gibbs steady-state distribution in the continuous amplitude space, realizing the functional extension from "only optimization" to "controllable sampling / generation", and this effect is directly generated by the natural laws constrained by optical nonlinearity and noise.

[0154] It also reduces manual parameter tuning, improves deployment efficiency and long-term stability: The training controller 30 updates the physical configuration of pumps, coupling, and external fields in real time based on measurement statistics within the device, reducing external offline trial and error. Combined with constraint projection and other operations, it improves stability and resists device drift.

[0155] On the hardware side, it can also improve feedback link efficiency and reduce latency / power consumption: by using dedicated MAC or cross-connect arrays and statistical calculations with fixed data paths, the involvement of general-purpose processors is reduced, thus lowering data handling and storage overhead. Achieving lower latency and better energy efficiency at the same sampling throughput represents a technical improvement in the configuration and function of computing devices.

[0156] Enhanced programmability and scalability: The programmable coupling matrix and external field injection structure support various topologies such as fully connected, sparse, and block-based. Noise injection and temperature scheduling enhance exploration capabilities. Support for rapid switching between multiple scenarios improves hardware versatility.

[0157] The third embodiment of this application provides a non-transitory machine-readable medium storing computer instructions for causing a computer to execute the training method of the coherent Ising machine 10 provided in the first embodiment of this application.

[0158] In the context of embodiments of this application, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable signal medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0159] It should be noted that the term "comprising" and its variations used in the embodiments of this application are open-ended, meaning "including but not limited to". The term "based on" means "at least partially based on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; and the term "some embodiments" means "at least some embodiments". The modifications of "one" and "multiple" mentioned in the embodiments of this application are illustrative and not restrictive. Those skilled in the art should understand that, unless explicitly indicated otherwise in the context, they should be understood as "one or more".

[0160] The user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in the embodiments of this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0161] The steps described in the method embodiments provided in this application can be performed in different orders and / or in parallel. Furthermore, the method embodiments may include additional steps and / or omit the steps shown. The scope of protection of this application is not limited in this respect.

[0162] The term "embodiment" in this specification refers to a specific feature, structure, or characteristic described in connection with an embodiment that may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily imply the same embodiment, nor does it imply independence from or alternative to other embodiments. The various embodiments in this specification are described in a related manner, with reference to each other for similar or identical parts. In particular, for apparatus, device, and system embodiments, since they are substantially similar to method embodiments, the description is relatively simple, and relevant details are referred to in the description of the method embodiments.

[0163] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of patent protection. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the appended claims.

Claims

1. A training method for a coherent Ising machine, characterized in that, The coherent Ising machine (10) includes a laser pulse generator (11) for generating N optical pulses with continuous amplitudes, where N is a natural number, and the N continuous amplitudes constitute an N-dimensional continuous amplitude vector. The training method includes: Noise is injected into the coherent Ising machine (10) so that, under the combined action of the physical parameters of the coherent Ising machine (10) and the noise, the continuous amplitude vector evolves to a steady-state distribution that follows the Gibbs distribution according to the Langevin dynamics evolution rule. The physical parameters include pump parameters, saturation parameters, coupling coefficients, bias field and gain. After taking the continuous amplitude vector collected under the current steady-state distribution as negative phase amplitude data, one or more of the physical parameters are updated based on the update amount obtained from the deviation between the negative phase amplitude data and the positive phase amplitude data, so that the continuous amplitude vector reaches the target steady-state distribution. The positive phase amplitude data comes from a pre-set training dataset. The step of obtaining the update amount based on the deviation between the negative phase amplitude data and the positive phase amplitude data includes: After calculating the first statistic of the negative phase amplitude data and the second statistic of the positive phase amplitude data respectively, the update quantity is generated based on the difference between the first statistic and the second statistic, wherein the first statistic and the second statistic each include one or more of the following statistics: The first moment, which is a statistic required to update the bias field, represents the amplitude of the i-th optical pulse. The mean of i∈N; The second moment, which is a statistic required to update the pump parameters, represents the amplitude of the i-th optical pulse. The average of the squares; The second-order correlation moment, which is a statistic required to update the coupling coefficient, represents the amplitude of the i-th optical pulse. and the amplitude of the j-th optical pulse The average of the products between them, j∈N; The fourth moment, which is a statistic required to update the saturation parameter, represents the average value of the fourth power of the amplitude of the i-th optical pulse; The amplitude energy correlation term is a statistic required to update the gain. The amplitude energy correlation term is the average energy of the coupling term and the bias term. The coupling term is the total energy of the interaction between all pulse pairs formed by the i-th optical pulse and the j-th optical pulse. The bias term is the energy of the i-th optical pulse under the action of the bias field.

2. The method according to claim 1, characterized in that, The update amount includes the update amount used to update the bias field amount. Then, the step of generating the update quantity based on the difference between the first statistic and the second statistic includes: Obtain the first moment of the first statistic The first moment of the second statistic The intensity D of the noise and the learning rate required to update the bias field quantity. ; In calculation and After calculating the difference between them, the obtained difference is compared with the gain and the learning rate. The product of the three factors, including the reciprocal of D, is added to the bias field before the update to obtain the update amount. .

3. The method according to claim 1, characterized in that, The update amount also includes an update amount for updating the pump parameters. Then, the step of generating the update quantity based on the difference between the first statistic and the second statistic further includes: Obtain the second moment of the first statistic The second moment of the second statistic The noise intensity D and the learning rate required to update the pump parameters. ; In calculation After calculating the difference between them, the obtained difference is then compared with the learning rate. The product of the inverse of 2D and the pump parameter before the update is added to obtain the update amount. .

4. The method according to claim 1, characterized in that, The update amount also includes an update amount for updating the coupling coefficient. Then, the step of generating the update quantity based on the difference between the first statistic and the second statistic further includes: Obtain the second-order correlation moments of the first statistic The second correlation moments of the second statistic The noise intensity D and the learning rate required to update the coupling coefficients. ; In calculation and After calculating the difference between them, the obtained difference is compared with the learning rate. The product of the inverse of 2D and the gain is added to the coupling coefficient before the update to obtain the update amount. .

5. The method according to claim 4, characterized in that, The method further includes: The coupling coefficient between all the pulse pairs is set. The coupling matrix J formed is symmetric and When =0, based on the update amount After updating the coupling coefficient, a symmetric projection operation and a zero diagonal constraint operation are performed on the coupling coefficient. The symmetric projection operation is the operation of projecting the coupling matrix J onto a symmetric matrix space, and the zero diagonal constraint operation is the operation of setting the elements on the diagonal of the coupling matrix J to zero.

6. The method according to claim 1, characterized in that, The update amount also includes an update amount for updating the saturation parameter. Then, the step of generating the update quantity based on the difference between the first statistic and the second statistic further includes: Obtain the fourth moment of the first statistic The fourth moment of the second statistic The noise intensity D and the learning rate required to update the saturation parameter. ; In calculation and After calculating the difference between them, the obtained difference is compared with the learning rate. The product of the inverse of 4D and the saturation parameter before the update is added to obtain the update amount. .

7. The method according to claim 1, characterized in that, The update amount also includes an update amount for updating the gain. Then, the step of generating the update quantity based on the difference between the first statistic and the second statistic further includes: Obtain the amplitude energy related term of the first statistic. The amplitude energy related term of the second statistic The intensity D of the noise and the learning rate required to update the gain. ,in, Let be the coupling coefficient between the i-th optical pulse and the j-th optical pulse. For coupling terms, For bias terms; In calculation and After calculating the difference between them, the obtained difference is compared with the reciprocal of D and the learning rate. The product of the two is added to the gain before the update to obtain the update amount. .

8. The method according to claim 7, characterized in that, The method further includes: Through the update amount During the update of the gain, the magnitude of the gain change is constrained based on set constraint rules. These constraint rules include magnitude constraints or regularization terms. The magnitude constraints limit the range of values ​​for the gain, and the regularization terms are used to calculate the update amount. A penalty term is added during the process to suppress excessive changes in the gain.

9. The method according to claim 1, characterized in that, The steps for updating the physical parameters based on the update amount include: The update amount used for the pump parameters The update is sent to the pump modulation unit (121) of the coherent Ising machine (10) so that the pump modulation unit (121) can update the amount of data. Modulate the optical pulse; and / or, The update amount used to update the saturation parameter The update is sent to the saturation control unit (122) of the coherent Ising machine (10) so that the saturation control unit (122) can update the amount of data. Control the saturation parameter within a stable range; and / or, The update amount used to update the gain Update the amount of the coupling coefficient. And the update amount of the bias field quantity. The data is sent to the feedback module (13) of the coherent Ising machine (10) so that the feedback module (13) updates the data based on the continuous amplitude under the current steady-state distribution of the optical pulse and the update amount. The update amount and the update amount The obtained feedback quantity is used to adjust the optical pulse, and the feedback quantity includes the total amount of the coupling effect of each optical pulse on other optical pulses and the amount of the bias field it receives.

10. The method according to claim 9, characterized in that, The method further includes: The negative phase amplitude data and the positive phase amplitude data are input to the calculation unit (132) of the feedback module (13) to calculate the first statistic and the second statistic through the calculation unit (132), and the update quantity is generated by the calculation unit (132) based on the difference between the first statistic and the second statistic.

11. A computing device based on a coherent Ising machine, characterized in that, The computing device based on the coherent Ising machine (10) is used to execute the training method of the coherent Ising machine (10) according to any one of claims 1 to 10, the device comprising: The coherent Ising machine (10) includes a laser pulse generator (11), a control module (12), and a feedback module (13). The control module (12) and the feedback module (13) are respectively connected to the laser pulse generator (11). The laser pulse generator is used to generate N optical pulses with continuous amplitude. The control module (12) is used to control the pump parameters and saturation parameters of the optical pulses. The feedback module (13) is used to adjust the optical pulses based on the feedback values ​​obtained from the continuous amplitude, gain, coupling coefficient, and bias field of the optical pulses. A noise controllable injector (20) is connected to the laser pulse generator (11) and the feedback module (13) respectively. The noise controllable injector (20) is used to generate controllable noise and inject the noise into the laser pulse generator (11) or the feedback module (13). Training controller (30) is connected to the control module (12) and the feedback module (13). The training controller (30) is used to obtain an update amount based on the deviation between the collected negative phase amplitude data and positive phase amplitude data, and to send the update amount to the control module (12) and the feedback module (13).

12. The computing device based on the coherent Ising machine according to claim 11, characterized in that, The update amount includes the update amount. Update volume Update volume Update volume and update volume The control module (12) includes: A pump modulation unit (121) is used to modulate the pump parameters; A saturation control unit (122) is used to control the saturation parameter within a stable range. The saturation control unit (122) and the pump modulation unit (121) are both connected to the training controller (30). The training controller (30) will update the amount The update amount is sent to the pump modulation unit (121). The update amount is sent to the saturation control unit (122). The update amount and the update amount The data is sent to the feedback module (13).

13. The computing device based on the coherent Ising machine according to claim 11 or 12, characterized in that, The coherent Ising machine (10) further includes a measurement unit (131) for measuring the continuous amplitude of the light pulse, and the feedback module (13) includes: The calculation unit (132) is connected to the measurement unit (131) and the training controller (30) respectively. The calculation unit (132) generates the feedback quantity after receiving the continuous amplitude. Feedback injection unit (133) is connected to the calculation unit (132) and the laser pulse generator (11). The feedback injection unit (133) is used to inject the feedback amount into the laser pulse generator (11) to adjust the light pulse. The training controller (30) can calculate the first statistic and the second statistic through the calculation unit (132), and generate the update quantity based on the difference between the first statistic and the second statistic through the calculation unit (132).

14. The computing device based on the coherent Ising machine according to claim 13, characterized in that, The feedback module (13) also includes: A sampling unit (134) is connected between the measurement unit (131) and the calculation unit (132). The sampling unit (134) is used to send amplitude data collected from the continuous amplitude to the calculation unit (132) during the evolution of the coherent Ising machine (10) under the current physical parameters and noise.

15. The computing device based on the coherent Ising machine according to claim 13, characterized in that, The computing unit (132) includes: The control chip is connected to the measurement unit (131) and the training controller (30) respectively. The control chip includes either a field-programmable gate array (FPGA) or an application-specific integrated circuit (ASIC). A computing device connected to the control chip, the computing device comprising either a multiplication and accumulation MAC array or a cross array.

16. A non-transitory machine-readable medium storing computer instructions, characterized in that, The computer instructions are used to cause the computer to perform the method according to any one of claims 1 to 10.

Images