JP2019036207A - Optimizing device, optimization method and optimization program - Google Patents

Abstract

To facilitate the optimization of an adversarial net.SOLUTION: An optimizing device according to the present application optimizes the parameters of a generator that generates new information so that a difference between previously generated information and correct answer information is reduced and a difference extractor for outputting a difference between the new information generated by the generator and the correct answer information, the difference being the one that serves as an index when the generator further generates new information. The optimizing device causes a quantum computation device that fits the parameters to the lattice points of a lattice model, sets the characteristic function of the lattice model on the basis of the learning policies of the generator and the difference extractor and finds using quantum fluctuation a ground state of the lattice model where the value of the characteristic function is minimum, to calculate the value of lattice points in the generated ground state of the characteristic function and output the values of the parameters that correspond to the values of the lattice points.SELECTED DRAWING: Figure 1

Description

  The present invention relates to an optimization device, an optimization method, and an optimization program.

  In recent years, Adversarial Nets technology has been proposed in which a plurality of learning devices such as SVM (Support Vector Machine) and DNN (Deep Neural Network) can be learned and mutually controlled. .

JP-T-2006-531343 gazette

“Generative Adversarial Nets”, Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, Yoshua Bengio, Internet <https://arxiv.org/pdf/1406.2661. pdf>, search on August 4, 2017 “Quantum Boltzmann Machine”, Mohammad H. Amin, Evgeny Andriyash, Jason Rolfe, Bohdan Kulchytskyy, Roger Melko, Internet <https://arxiv.org/pdf/1601.02036.pdf>, search August 4, 2017

  Here, it takes a lot of time and computational resources to learn features and optimize various learners, so further optimization of Adversal Real Net using multiple learners There is a tendency to require a lot of time and computational resources.

  The present application has been made in view of the above, and an object thereof is to facilitate optimization of an adversarial net.

  The optimization apparatus according to the present application includes a generator that generates new information so that a difference between previously generated information and correct information is reduced, and a difference between the information newly generated by the generator and the correct information. An optimization device for optimizing a parameter with a difference extractor that outputs a difference as an index when the generator further generates new information, wherein the parameter is used as a lattice point of a lattice model. Fitting, a quantum function that sets a characteristic function of the lattice model based on a learning policy between the generator and the difference extractor, and uses a quantum fluctuation to obtain a ground state of the lattice model that minimizes the value of the characteristic function The calculation device is configured to calculate a value of a lattice point in a ground state of the generated characteristic function, and to output a value of the parameter corresponding to the value of the lattice point.

  According to one aspect of the embodiment, optimization of the adversarial net can be facilitated.

FIG. 1 is a diagram illustrating an example of optimization processing according to the embodiment. FIG. 2 is a diagram illustrating an example of a functional configuration of the optimization apparatus according to the embodiment. FIG. 3 is a diagram for explaining an example of processing executed by the quantum computing device according to the embodiment. FIG. 4 is a flowchart for explaining an example of processing executed by the optimization apparatus according to the embodiment.

  Hereinafter, a mode for carrying out an optimization apparatus, an optimization method, and an optimization program according to the present application (hereinafter referred to as “embodiment”) will be described in detail with reference to the drawings. Note that the optimization apparatus, the optimization method, and the optimization program according to the present application are not limited by this embodiment. In the following embodiments, the same portions are denoted by the same reference numerals, and redundant description is omitted.

[1. Optimization process)
First, an example of the optimization process according to the embodiment will be described with reference to FIG. FIG. 1 is a diagram illustrating an example of optimization processing according to the embodiment. In FIG. 1, an example of an optimization process for optimizing a learner parameter called an adversarial net will be described.

  First, an adversary real network AN optimized by the optimization device 10 will be described. For example, the adversarial net includes a content generator (hereinafter sometimes referred to as “G”) that generates arbitrary content such as text and images, the content generated by the content generator, and predetermined correct answer data. And a difference extractor (hereinafter, may be described as “D”).

  Here, the “difference” extracted by the difference extractor is not simply a numerical difference between the content and the correct answer data but a semantic difference between the content and the correct answer data. As a more specific example, the content generator generates an image reflecting the style of a painter. On the other hand, the difference extractor uses the picture of the painting actually created by the painter as correct data, and compares the style reflected in the image generated by the content generator with the style reflected in the correct data. The difference extractor extracts a style difference, that is, a semantic difference between the image generated by the content generator and the image that is the correct data, and outputs the extracted difference. Note that the content and correct data generated by the content generator can adopt not only images and texts but also arbitrary content such as moving images and voices.

  In such an adversa real net, in order to bring the content generated by the content generator closer to correct data, the content generator is modified so that the difference extracted by the difference extractor becomes smaller, and the content generator What is necessary is just to learn a difference extractor so that the difference of the produced | generated content and correct data may be extracted more correctly. That is, in order to learn the features of correct data in the Adversal Real Net, the difference extractor learns so that the difference extracted by the difference extractor becomes larger, and the difference extracted by the difference extractor becomes smaller. In this way, the content generator may be learned. By learning the content generator and the difference extractor based on such a learning policy, it is possible to learn the features of correct data with high accuracy.

  However, there are cases where a large amount of correct answer data and computational resources are required for learning such an adversarial net. For example, the content generator and difference extractor of Adversal Real Net is a model in which multiple nodes are connected in multiple stages, such as DNN (Deep Neural Network), and the weight (connection for connection) between each node. It is composed of a model that learns an arbitrary feature of data by correcting the value of the coefficient). However, in order to make such a model learn a feature, using a learning method such as back propagation, Since the connection coefficient is corrected in accordance with the learning policy described above, there is a possibility that computational resources required for learning increase.

  Therefore, the optimization apparatus 10 uses the quantum computation such as the machine learning quantum annealing as shown in Non-Patent Document 2 to improve the optimization of the adversarial net. More specifically, the optimization apparatus 10 includes a content generator that generates new information so that a difference between the generated information and the correct answer information is reduced, and information and correct information that are newly generated by the content generator. The parameter is optimized with a difference extractor that outputs a difference that serves as an index when the content generator generates new information. For example, the optimization apparatus 10 uses the connection coefficient between nodes of the content generator and the difference extractor and the connection relationship between the nodes as parameters of the adversarial net, and sets the parameters as lattice points of the lattice model (Ising model) Fitting, a characteristic function (Hamiltonian) of the lattice model is set based on the learning policy between the content generator and the difference extractor. Then, the optimization apparatus 10 causes the quantum calculation apparatus that obtains the ground state of the lattice model having the minimum characteristic function value using the quantum fluctuation to calculate the value of the lattice point in the ground state of the generated characteristic function. The value of the calculated grid point is output.

  In other words, the optimization apparatus 10 sets a system in which the values of the parameters of the adversarial net, that is, the parameters of the content generator and the difference extractor are associated with each lattice point of the Ising model, and the content is set for the system. Set a characteristic function that reflects the learning policy of the generator and difference extractor. The optimization apparatus 10 uses a technique such as quantum annealing for machine learning as shown in Non-Patent Document 2 to converge the state of the lattice model to the ground state where the value of the characteristic function is the minimum value. Then, an optimum value corresponding to the value of the characteristic function, that is, an optimum solution of the parameters of the adversarial net for the learning policy of the content generator and the difference extractor is obtained.

  Thus, since the optimization apparatus 10 calculates the parameters of the content generator and the difference extractor by quantum calculation, it is possible to reduce the calculation resources required for learning the adversarial net.

[2. (Fitting of learning policy)
Hereinafter, an example of processing in which the optimization apparatus applies the learning policy of the content generator or the difference extractor to the characteristic function will be described. In the following description, an example of processing for mapping an adversal real net to an Ising model will be described. A Hamiltonian in which an Adversal Realnet is mapped to a model may be generated. For example, consider a one-dimensional Ising model in which α lattices are arranged in a line. The state function of such an Ising model can be expressed by the following equation (1).

  Here, an Ising model in which two grids are arranged is considered. If each lattice takes a value of 0 or 1 at the time of observation, each lattice is superposed with 0 and 1 at the time of quantum calculation. For this reason, in the whole Ising model, as shown by the following formula (2), a superposed state of four states occurs at the time of quantum calculation.

  Here, the temporal change of the whole Ising model is considered. For example, considering the time differentiation of equation (2), the following equation (3) can be obtained.

Here, H (t) shown in the equation (3) is an operator that converts the state function of the system, that is, a Hamiltonian, and can be expressed by the following equation (4). Here, A (t) and B (t) are coefficients that change with time. Σ ^x _i is a Pauli matrix shown in the following equation (5). H ₀ is a Hamiltonian represented by the following formula (6).

  Here, consider the time evolution of the Hamiltonian shown in Equation (4). For example, the initial state H (t = 0) of the Hamiltonian shown in the equation (4) can be expressed by the following equation (7).

  Here, when both A (t) and B (t) shown in Equation (4) take positive values, A (t) increases the Hamiltonian shown in Equation (7) with time, and B (t) Will decrease the Hamiltonian represented by equation (7) with time. That is, A (t) increases the energy of the system with time evolution, and B (t) decreases the energy of the system with time evolution. In a technique called quantum annealing, the problem to be calculated is associated with such time evolution, and the time evolution of the system is controlled according to the content of the problem, so that the final state of the system is the solution to the problem. .

  Here, let us consider a process in which the learning policy of the content generator and the difference extractor constituting the adversal real net is made to correspond to the Hamiltonian of the Ising model. For example, an adversarial network composed of a content generator and a difference extractor can be regarded as a Boltzmann machine. There has been proposed a method of calculating a parameter of a quantum Boltzmann machine obtained by quantizing such a Boltzmann machine by quantum annealing (see, for example, Non-Patent Document 2).

  For example, the Hamiltonian of the Boltzmann machine can be expressed by the above-described equation (4), and the process for obtaining the minimum value of the function represented by the expression (4) corresponds to the process for calculating the Boltzmann machine parameter. Here, the energy function of the Boltzmann machine can be expressed by the following formula (8), and the process for obtaining the minimum value of the energy function represented by the formula (8) also corresponds to the process for calculating the parameters of the Boltzmann machine. X in Equation (8) is the state of the Boltzmann machine, and θ is a parameter for minimizing an error from the true model distribution.

  Here, the conditional probability P (X | θ) between the state X and the parameter θ can be expressed by the following equation (9). Further, Z (θ) in the formula (9) can be expressed by the following formula (10).

The optimization apparatus 10 regards the adversal real net as a quantum Boltzmann machine, and executes processing for calculating the parameters of the quantum Boltzmann machine, so that the parameters of the content generator and the difference extractor included in the adversal net Is calculated. For example, correct data is p _data (x), and content generated by the content generator is p _g (x). In such a case, the learning policy of the adversarial net can be expressed by the following equation (11) (for example, see Non-Patent Document 1). Here, D in Expression (11) is a function corresponding to the difference extractor, and G is a function corresponding to the content generator.

  That is, in the Adversal Real Net, the content generator is trained to generate content that minimizes the difference extracted by the difference extractor, and the difference between the correct data and the content generated by the content generator The difference extractor is learned so that becomes larger. That is, in the adversarial net, the parameters of the content generator are set so as to generate content that minimizes the difference extracted by the difference extractor, and the correct data and the content generated by the content generator are set. The parameter of the difference extractor is set so that the difference between the two becomes larger.

  The learning policy of such an adversarial net is that Is (T) shown in Equation (4) increases the energy of the system with time development, and B (t) decreases the energy of the system with time development. It can be mapped to the time evolution policy of the model Hamiltonian. Therefore, the optimization apparatus 10 regards the adversarial net as a quantized Boltzmann machine, and sets the Hamiltonian of the quantized Boltzmann machine as a Hamiltonian. More specifically, the optimization device 10 regards the content generator learning policy of minimizing the difference output from the difference extractor as a policy of minimizing the value of the term that reduces the value of the Hamiltonian. The Hamiltonian is set by regarding the learning policy of the difference generator that maximizes the difference between the content newly generated by the content generator and the correct data as the policy of maximizing the value of the Hamiltonian.

For example, the Hamiltonian parameter of the quantum Boltzmann machine can be expressed by the following equation (12). Incidentally, _{H z} of formula (12), equation (4), corresponding to _{H 0} shown in (6).

  Here, the second term of the equation (12) serves to maximize the energy of the entire system. Therefore, the optimization apparatus 10 inverts the sign of the second term as shown in Expression (13). That is, the optimization apparatus 10 inverts the sign of the term indicating the interaction between the nearest lattices in the Ising model.

  The whole of the equation (13) works to maximize the energy of the entire system. Therefore, the optimization apparatus 10 maps the learning policy of the difference extractor to the entire Hamiltonian represented by Expression (13). In addition, the second term of equation (13) serves to minimize the energy of the system. Therefore, the optimization apparatus 10 maps the learning policy of the content generator to the second term of Expression (13).

  By setting a Hamiltonian that maps the learning policy in this way and introducing the Ising model to the ground state where the energy of the system is minimized by the quantum annealing method that assumes machine learning as in Non-Patent Document 2, the Ising model is derived. Each lattice point of the model converges to a value according to the learning policy of the adversarial net. That is, each lattice point converges to the value of each parameter of the Adversal Realnet that reflects the set learning policy. As a result, the optimization apparatus 10 can reduce the calculation cost required for learning the adversarial net.

[3. Example of processing)
Next, an example of optimization processing executed by the optimization device 10 will be described with reference to FIG. First, the optimization apparatus 10 receives initial parameters and correct data of the adversarial network AN (step S1). Here, the initial parameters are parameters of the content generator and the difference extractor.

  In such a case, the optimization apparatus 10 applies the parameters of the content generator and the difference extractor to the lattice points, and sets a Hamiltonian for the lattice model based on the learning policy between the content generator and the difference extractor. (Step S2). For example, the optimization apparatus 10 regards the difference extractor learning policy of maximizing the difference between the correct answer data and the content generated by the content generator as a process of maximizing the entire expression (13). The content generator learning policy of minimizing the difference output from the difference extractor is regarded as a process of minimizing the second term of Expression (13), and the Hamiltonian is set.

Then, the optimization apparatus 10 searches for an optimal solution using quantum computation (step S3). For example, the quantum computing device uses the quantum annealing for machine learning as described in Non-Patent Document 2 to develop the time from the superposition state where the values of each qubit are uniform to the Ising model indicated by the generated Hamiltonian. Then, a parameter list indicating the values of the parameter values ω _{11 to} ω _jj mapped to each lattice point is output (step S4).

  Note that the process of searching the ground state of the Ising model generated by the optimization device 10 is not limited to the above description. For example, in quantum computation, an adiabatic model such as a lattice model and a circuit model using an arbitrary quantum circuit are equivalent. For this reason, since quantum computation using quantum annealing is equivalent to a quantum circuit that continuously operates a quantum gate, the optimization apparatus 10 uses an Ising model represented by the generated Hamiltonian as a quantum circuit that combines quantum gates. Can be optimized by any hardware that reproduces. For example, the optimization apparatus 10 may search the ground state of the generated Ising model using a quantum calculation apparatus that performs quantum calculation using nuclear magnetic resonance, quantum dots, Josephson elements, ion traps, photons, and the like. .

[4. Optimization device configuration)
Next, the configuration of the optimization apparatus 10 according to the embodiment will be described with reference to FIG. FIG. 2 is a diagram illustrating an example of a functional configuration of the optimization apparatus according to the embodiment. As shown in FIG. 2, the optimization device 10 is connected to an input device 20 and an output device 30. The optimization apparatus 10 includes an acquisition unit 11, a generation unit 12, a quantum calculation device 13, and an output unit 17. The quantum computation device 13 includes a state reproduction unit 14, an operation unit 15, and an observation unit 16. The state reproduction unit 14 includes a plurality of quantum bits 14a to 14d.

  The input device 20 is an input device for inputting to the optimizing device 10, and is, for example, a reading device that can read data from a recording medium such as an HDD (Hard Disk Drive) or a flash memory. For example, when the input device 20 reads the parameters of the adversarial net AN from a recording medium such as a flash memory, the input device 20 outputs the parameters of the adversarial net AN to the optimization device 10.

  The output device 30 is realized by, for example, a monitor, a printer, or the like, and receives an adversarial net parameter optimized by the optimization device 10, that is, a parameter list indicating a coupling coefficient between the content generator and the difference extractor. The output device displays or prints the received parameter list.

  When the acquisition unit 11 acquires each parameter of the Adversal Real Network AN, the acquisition unit 11 outputs each parameter to the generation unit 12.

  When the generating unit 12 acquires each parameter of the adversarial net AN, it generates a characteristic function for calculating the parameter value of the adversarial net AN, that is, a Hamiltonian according to a predetermined learning policy. More specifically, the generation unit 12 applies parameters to lattice points of the lattice model. The generation unit 12 is a content generator that generates new information so that a difference between the generated information and the correct answer information is reduced, and a difference between the information newly generated by the content generator and the correct information. The characteristic function for optimizing the parameter with the difference extractor that outputs the difference that becomes an index when the content generator generates further new information is set based on the learning policy of the adversary net AN To do.

  For example, the generation unit 12 regards a set of a content generator and a difference extractor as a quantized Boltzmann machine, and sets a quantized Boltzmann machine Hamiltonian as a characteristic function. Further, the generation unit 12 considers the learning policy of the content generator that minimizes the difference output from the difference extractor as the policy of minimizing the value of the term that decreases the value of the characteristic function. Further, the generation unit 12 considers the learning policy of the difference generator that maximizes the difference between the information newly generated by the content generator and the correct answer information as the policy that maximizes the value of the characteristic function. Then, the generation unit 12 sets a characteristic function. More specifically, the generation unit 12 inverts the sign of the term indicating the interaction between the nearest grids in the characteristic function, and assigns the learning policy of the content generator to the term whose sign is inverted.

  Then, the generation unit 12 outputs a correspondence between each lattice of the Ising model and each parameter and a characteristic function to the quantum calculation device 13.

  The quantum computing device 13 is a device that searches for the ground state of the Ising model generated by the generating unit 12 using quantum computation, and is a so-called quantum computer. For example, the quantum computing device 13 has a plurality of qubits 14a to 14d that can hold a plurality of values in a superposed state, and can reproduce the state of the Ising model using each qubit 14a to 14d. A reproduction unit 14 is included.

  When the operation unit 15 of the quantum computing device 13 accepts the characteristic function generated by the generation unit 12, the state reproduction unit 14 includes the qubits 14a to 14d and the coupling coefficient between the qubits 14a to 14d. , The ground state of the Ising model indicated by the characteristic function received from the generation unit 12 is calculated quantitatively. That is, the operation unit 15 controls each qubit 14a to 14d so as to calculate each parameter of the adversarial net using machine learning quantum annealing as in Non-Patent Document 2. And the observation part 16 observes the value of each quantum bit 14a-14d of the state reproduction part 13 fully developed in time. In other words, the quantum computing device 13 obtains the ground state of the lattice model in which the value of the characteristic function becomes the minimum value using the quantum fluctuation, and calculates the value of the lattice point in the ground state of the characteristic function.

  Here, the concept of processing executed by the quantum computation device 13 will be described with reference to FIG. FIG. 3 is a diagram for explaining an example of processing executed by the quantum computing device according to the embodiment. In the example shown in FIG. 3, the characteristic function, that is, the value of the Hamiltonian is set on the vertical axis, and the combination of the values of the grid points is set on the horizontal axis. Each combination of values is indicated by a solid line.

  For example, the quantum computing device 13 sets the superposed state of values that can take the values of the quantum bits 14a to 14d. As a result, the quantum computation device 13 reproduces all combinations of the values of the respective lattice points in the Ising model as indicated by the dotted circles in FIG. Then, as indicated by the dotted arrows in FIG. 3, the quantum computing device 13 develops the state of each of the qubits 14 a to 14 d to the state indicated by the Hamiltonian generated by the generation unit 12, thereby Reproduce the Hamiltonian value for each combination of values. Such processing is performed while the qubits 14a to 14d are kept in the quantum state.

  Here, the state in which the value of the Hamiltonian takes the minimum value is the most stable state in the system. For this reason, when the qubits 14a to 14d that are sufficiently developed in time with the quantum state maintained are observed, there is a possibility that a state ω in which the value of the Hamiltonian is minimum is observed as shown in FIG. Is expensive. Therefore, the quantum computing device 13 observes the value of each of the qubits 14a to 14d that has been sufficiently developed while maintaining the quantum state, so that the value of the Hamiltonian is minimized without falling into a minimum value. Specify the value of each grid point.

  Returning to FIG. 2, the description will be continued. When the output unit 17 receives the values of the qubits 14 a to 14 d observed by the observation unit 16, the output unit 17 outputs a parameter list indicating the optimized parameters to the output device 30 based on the values. For example, the output unit 17 sets the value of each qubit 14a to 14d as the value of the lattice point of the Ising model corresponding to each qubit 14a to 14d. Then, the output unit 17 acquires the optimized parameter by inverse mapping the parameter from each lattice point of the Ising model, and outputs a parameter list including the acquired parameter to the output device 30.

[5. Flow of processing executed by optimization device]
Next, the flow of processing executed by the optimization apparatus 10 according to the embodiment will be described with reference to FIG. FIG. 4 is a flowchart for explaining an example of processing executed by the optimization apparatus according to the embodiment. First, the optimization apparatus 10 accepts parameters of the adversarial net (step S101). Subsequently, the optimization apparatus 10 maps each parameter of the Adversal Real Net to each lattice point of the Ising model (Step S102), and based on the learning policy of the content generator and the learning policy of the difference extractor, Boltzmann The Hamiltonian of the machine is set (step S103). Subsequently, the optimization apparatus 10 uses the set Hamiltonian to temporally develop the Ising model into the ground state (step S104) and observe the ground state (step S105). Then, the optimization apparatus 10 outputs the observation result as the optimization result, that is, the value of each parameter of the optimized adversarial net (step S106), and ends the process.

[6. (Modification)
The optimization apparatus 10 according to the above-described embodiment may be implemented in various different forms other than the above-described embodiment. Therefore, in the following, another embodiment of the optimization device 10 will be described.

[6-1. Others]
Here, the optimization apparatus 10 may perform processing in consideration of a route that is not used when the Adversal Real Net is actually operated. For example, the optimization apparatus 10 may calculate an optimal parameter set by quantum calculation using the coupling coefficient set for all connection paths including the dummy path as a parameter.

  Further, the optimization apparatus 10 may perform processing in consideration of convolution. For example, the optimizing device 10 may put a combination of coupling coefficients that are convolved between different nodes into a superposed state as a parameter. That is, the optimization apparatus 10 obtains combinations of coupling coefficient values set for all connection paths for each convolution mode, and obtains all combinations of combinations of coupling coefficients obtained for all convolution modes. You may be in a superposition state.

[6-2. (Hardware configuration)
The optimization device 10 described above has the quantum calculation device 13 that searches the ground state of the Ising model indicated by the Hamiltonian. However, the embodiment is not limited to this. For example, the optimization apparatus 10 may include only the acquisition unit 11, the generation unit 12, and the output unit 17, and may cause the quantum computation device 13 installed outside to search for a ground state.

  Among the functional configurations of the optimization device 10, the processing performed by the acquisition unit 11, the generation unit 12, and the output unit 17 is an integrated circuit such as an ASIC (Application Specific Integrated Circuit) or an FPGA (Field Programmable Gate Array). May be realized. The processing performed by the acquisition unit 11, the generation unit 12, and the output unit 17 is stored in a storage device inside the optimization device 10 by a processor such as a CPU (Central Processing Unit) or an MPU (Micro Processing Unit), for example. The optimization program may be realized by executing a RAM (Random Access Memory) as a work area.

[7. effect〕
As described above, the optimization device 10 includes a content generator that generates new information so that the difference between the previously generated information and the correct answer data is small, and the information and correct answer data that are newly generated by the content generator. The optimization of the parameters with the difference extractor that outputs the difference, which is an index when the content generator generates new information, is performed. Here, the optimization apparatus 10 applies the parameter to the lattice point of the lattice model, and sets the characteristic function of the lattice model based on the learning policy between the content generator and the difference extractor. Further, the optimization apparatus 10 causes the quantum calculation apparatus that obtains the ground state of the lattice model having the minimum characteristic function value using the quantum fluctuation to calculate the value of the lattice point in the ground state of the generated characteristic function. . Then, the optimization apparatus 10 outputs a parameter value corresponding to the value of the grid point.

  As described above, the optimization device 10 uses the quantum calculation to calculate the parameters of the content generator and the difference extractor included in the adversarial net, so that the calculation resources and calculation time required for learning the adversarial net are reduced. As a result of the reduction, optimization of the adversa real net can be facilitated.

  Further, the optimization apparatus 10 regards the set of the content generator and the difference extractor as a quantized Boltzmann machine, and sets the quantized Boltzmann machine Hamiltonian as a characteristic function. Further, the optimization device 10 regards the learning policy of the content generator that minimizes the difference output from the difference extractor as the policy that minimizes the value of the term that decreases the value of the characteristic function, and the content generator The characteristic function is set by considering the learning policy of the difference content generator that maximizes the difference between the newly generated information and the correct data as the policy that maximizes the value of the characteristic function. For example, the optimization apparatus 10 inverts the sign of the term indicating the interaction between the nearest lattices in the characteristic function, and applies the learning policy of the content generator to the term whose sign is inverted.

  As a result of such processing, the optimization apparatus 10 performs processing according to the learning policy of the Adversal Real in the Ising model and the processing of optimizing the parameters of the Adversal Net according to the learning policy of the Adversal Real. Thus, it can be put into the process of optimizing the value of the lattice. As a result, the optimization apparatus 10 can facilitate the optimization of the adversarial net.

  As described above, some of the embodiments of the present application have been described in detail with reference to the drawings. However, these are merely examples, and various modifications, including the aspects described in the disclosure section of the invention, based on the knowledge of those skilled in the art, It is possible to implement the present invention in other forms with improvements.

  Moreover, the above-mentioned “section (module, unit)” can be read as “means”, “circuit”, and the like. For example, the generation unit can be read as generation means or a generation circuit.

DESCRIPTION OF SYMBOLS 10 Optimization apparatus 11 Acquisition part 12 Generation part 13 Quantum calculation apparatus 14 State reproduction part 14a-14d Qubit 15 Operation part 16 Observation part 17 Output part 20 Input apparatus 30 Output apparatus

Claims (6)

A generator that generates new information so that a difference between the previously generated information and the correct answer information is reduced, and a difference between the information newly generated by the generator and the correct answer information, and the generator further includes: An optimization device for optimizing a parameter with a difference extractor that outputs a difference as an index when generating new information,
Applying the parameters to lattice points of the lattice model, and setting a characteristic function of the lattice model based on a learning policy between the generator and the difference extractor,
A quantum computing device that obtains a ground state of a lattice model having a minimum value of a characteristic function using quantum fluctuations, and calculates a value of a lattice point in the ground state of the generated characteristic function,
The optimization apparatus, wherein the parameter value corresponding to the value of the grid point is output. The optimization device includes:
The set of the generator and the difference extractor is regarded as a quantized Boltzmann machine, and a quantized Boltzmann machine Hamiltonian is set as the characteristic function. Optimization device. The optimization device includes:
Information that is newly generated by the generator, considering the learning policy of the generator that minimizes the difference output from the difference extractor as a policy that minimizes the value of the term that decreases the value of the characteristic function. The characteristic function is set by regarding the learning policy of the difference generator that maximizes the difference between the correct answer information and the correct answer information as a policy that maximizes the value of the characteristic function. The optimization device described in 1. The optimization device inverts the sign of a term indicating the interaction between nearest lattices in the characteristic function, and applies the learning policy of the generator to the term obtained by inverting the sign. Item 4. The optimization device according to Item 3. A generator that generates new information so that a difference between the previously generated information and the correct answer information is reduced, and a difference between the information newly generated by the generator and the correct answer information, and the generator further includes: An optimization method executed by an optimization device that optimizes a parameter with a difference extractor that outputs a difference serving as an index when generating new information,
Applying the parameters to lattice points of the lattice model, and setting a characteristic function of the lattice model based on a learning policy between the generator and the difference extractor,
A quantum computing device that obtains a ground state of a lattice model having a minimum value of a characteristic function using quantum fluctuations, and calculates a value of a lattice point in the ground state of the generated characteristic function,
An optimization method characterized by outputting the value of the parameter corresponding to the value of the grid point. A generator that generates new information so that a difference between the previously generated information and the correct answer information is reduced, and a difference between the information newly generated by the generator and the correct answer information, and the generator further includes: A computer having an optimization device for optimizing a parameter with a difference extractor that outputs a difference serving as an index when generating new information applies the parameter to a lattice point of a lattice model, and the generator and the difference Based on the learning policy with the extractor, set the characteristic function of the lattice model,
A quantum computing device that obtains a ground state of a lattice model having a minimum value of a characteristic function using quantum fluctuations, and calculates a value of a lattice point in the ground state of the generated characteristic function,
The optimization program for performing the process which outputs the value of the parameter corresponding to the value of the lattice point.

Images