Information processing program, information processing device, and information processing method
By applying translational symmetry and branching to QCNNs, the inefficiencies in calculation cost and time are addressed, enhancing the computational efficiency and enabling the processing of larger datasets with reduced resources.
Patent Information
- Application Number
- PCT/JP2023/019573
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2023-05-25
- Publication Date
- 2024-11-28
AI Technical Summary
Quantum convolutional neural networks (QCNNs) face inefficiencies in calculation cost and time as problem size increases, requiring numerous updates of variational parameters and inefficient measurement processes due to the need to execute quantum circuits multiple times to obtain a single output.
Implementing translational symmetry and branching in QCNNs to utilize discarded qubits for parallel computation, allowing for simultaneous measurement of multiple quantum states and reducing the number of required circuit executions, thereby enhancing measurement and execution efficiency.
This approach significantly improves the computational efficiency of QCNNs by reducing the time and resources needed for training and prediction tasks, enabling the processing of larger datasets with reduced computational costs.
Smart Images

Figure JP2023019573_28112024_PF_FP_ABST
Abstract
Description
Information processing program, information processing device, and information processing method
[0001] The present invention relates to machine learning technology using a quantum computer.
[0002] Machine learning methods using quantum computers are expected to be applied to a variety of fields, including quantum physics and finance, as they can process huge data sets that cannot be handled by classical computers, which are the general-purpose computers currently in use. In particular, quantum neural networks, which are neural networks constructed using quantum circuits, are attracting great expectations as a new technology that may surpass classical neural networks.
[0003] There is also a quantum convolutional neural network (QCNN), which is one type of quantum neural network and can classify quantum states. QCNN is an extension of convolutional neural networks (CNN), a representative classical machine learning model, to quantum machine learning, and is expected to be able to process larger data sets than classical CNNs.
[0004] Japanese Patent Publication No. 2019-096334 Japanese Patent Publication No. 2022-509841 U.S. Patent Application Publication No. 2022 / 0101164 U.S. Patent Application Publication No. 2022 / 0253741
[0005] However, QCNNs suffer from the problem of increasing computational costs as the problem size increases. For example, training a QCNN requires repeatedly updating a large number of variational parameters against a large amount of training data, which takes a very long time. Furthermore, unlike classical CNNs, quantum computing requires running a single quantum circuit multiple times to obtain a single output and calculating the average of the measurements. Furthermore, quantum computers are a developing technology, and in order to solve large-scale problems with the current limited computing resources, it is necessary to reduce computational costs as much as possible.
[0006] In one aspect, the objective is to make QCNN more efficient.
[0007] In one aspect, the information processing program causes a computer to execute the following process: inputting the quantum state of each of a plurality of quantum bits into a quantum circuit designed so that each layer of the QCNN is symmetric with respect to a translation operation of the quantum bit; performing an operation in a first layer of the QCNN on the plurality of quantum bits; dividing the plurality of quantum bits operated in the first layer into a plurality of quantum bit sets so as to be symmetric with respect to the translation operation; branching each of the plurality of quantum bits in the quantum bit set into a different branch of the quantum circuit to perform an operation in each layer of the QCNN; and measuring the quantum state of each of the plurality of quantum bits in the final state.
[0008] In one aspect, QCNN can be implemented more efficiently.
[0009] FIG. 1 is a diagram for explaining a quantum circuit of a QCNN. FIG. 2 is a diagram for explaining translational symmetrization and branching of a QCNN. FIG. 3 is a flowchart showing the overall processing flow including confirmation of prerequisites of this embodiment. FIG. 4 is a diagram for explaining the prerequisites of this embodiment. FIG. 5 is a diagram showing an example of a configuration of an information processing device 10 according to this embodiment. FIG. 6 is a diagram for explaining translational symmetrization of each layer of a QCNN according to this embodiment. FIG. 7 is a diagram showing an example of a translationally symmetric QCNN layer according to this embodiment. FIG. 8 is a diagram for explaining branching of a QCNN according to this embodiment. FIG. 9 is a diagram for explaining branching of n qubits according to this embodiment. FIG. 10 is a diagram for explaining branching of multiple qubits in a high-dimensional case according to this embodiment. FIG. 11 is a flowchart showing the flow of branching processing of a QCNN according to this embodiment. FIG. 12 is a diagram for explaining averaging of qubit measurement values according to this embodiment. FIG. 13 is a diagram (1) for explaining equivalence of qubit measurement values according to this embodiment. FIG. 14 is a diagram (2) for explaining equivalence of qubit measurement values according to this embodiment. FIG. 15 is a flowchart showing the flow of the process of measuring the output by a QCNN branched in translational symmetry according to this embodiment. FIG. 16 is a diagram for explaining a method of measuring the gradient of the QCNN output according to this embodiment. FIG. 17 is a diagram (1) for explaining a method of measuring the gradient when one parameter controls multiple gates according to this embodiment. FIG. 18 is a diagram (2) for explaining a method of measuring the gradient when one parameter controls multiple gates according to this embodiment. FIG. 19 is a diagram (1) for explaining the efficiency measurement of the gradient by a branched QCNN according to this embodiment. FIG. 20 is a diagram (2) for explaining the efficiency measurement of the gradient by a branched QCNN according to this embodiment. FIG. 21 is a diagram (3) for explaining the efficiency measurement of the gradient by a branched QCNN according to this embodiment. FIG. 22 is a flowchart showing the flow of the process of measuring the gradient of the QCNN output according to this embodiment.FIG. 23 is a flowchart showing the flow of the differential measurement process using the parameter shift method according to this embodiment. FIG. 24 is a diagram showing an example of a phase diagram of the ground state of a Hamiltonian. FIG. 25 is a diagram showing an example of the relationship between the phase diagram of the ground state of a Hamiltonian and training data for a QCNN. FIG. 26 is a diagram showing an example of a change in the loss function due to variational parameter updating. FIG. 27 is a diagram showing an example of prediction results for each problem size for training data using a trained QCNN. FIG. 28 is a diagram showing an example of prediction results for each problem size for unknown data using a trained QCNN. FIG. 29 is a diagram showing an example of the probability distribution of estimated values. FIG. 30 is a diagram showing an example of quantum circuits for a branched QCNN and an unbranched QCNN. FIG. 31 is a diagram showing an example of measurement efficiency due to branching of a QCNN for three representative states. FIG. 32 is a diagram showing another example of measurement efficiency due to branching of a QCNN for three representative states. FIG. 33 is a diagram showing an example of the hardware configuration of an information processing device 10 according to this embodiment.
[0010] Hereinafter, examples of the information processing program, information processing device, and information processing method according to the present embodiment will be described in detail with reference to the accompanying drawings. Note that the present embodiment is not limited to these examples. Furthermore, each example can be appropriately combined within a consistent range.
[0011] First, we will explain the quantum convolutional neural network (QCNN). The QCNN is a convolutional neural network (CNN), a representative classical machine learning model, that is constructed using quantum circuits and extended to quantum machine learning.
[0012] FIG. 1 is a diagram illustrating a quantum circuit of a QCNN. As shown in FIG. 1, the quantum circuit of a QCNN includes three layers: a C layer, a P layer, and an FC layer. The C layer is, for example, a convolution layer, which performs a local unitary transformation and extracts features. The P layer is, for example, a pooling layer, which reduces the number of qubits and performs coarse-graining of the quantum state. The FC layer is, for example, a fully connected layer, which performs a unitary transformation that connects all qubits together. In a quantum circuit of a QCNN, for example, by alternating between convolution and pooling, quantum correlation features of various length scales can be trained. In addition, in a quantum circuit of a QCNN, for example, the C layer and the P layer are alternately operated, and when the number of qubits has been sufficiently reduced, the fully connected layer is operated. Then, the quantum state of the qubits remaining after the fully connected layer is operated is measured, and the expected value is used as the output of the QCNN.
[0013] Also, in FIG. 1 ~U 5 The unitary transformation represented by is trained using training data. The training data is made up of, for example, pairs of quantum states and correct labels, and can be expressed by the following formula (1).
[0014]
[0015] In formula (1), for example, N t is the number of training data. The correct label value corresponds to the prediction of QCNN. For example, in image recognition, i =1" for cat image, "y i If "=0", the image is classified as anything other than a cat. The loss function during training can be expressed, for example, as the following equation (2).
[0016]
[0017] In equation (2), Z1 is, for example, the Z operator of the first quantum bit and the physical quantity to be measured. Furthermore, for example, U(θ) is the entire unitary transformation of the QCNN, and θ is the parameter of the unitary transformation. θ is then optimized to minimize the loss function expressed in equation (2). An example of an optimization method is the existing gradient descent method, which optimizes parameters based on the gradient of the loss function. In parameter optimization, for example, the gradient of the quantum computer loss function is calculated, and a classical computer updates the parameters based on the results. These processes are repeated alternately many times to optimize the parameters. Finally, the trained QCNN makes predictions for unknown data.
[0018] While QCNNs can process larger datasets than classical CNNs, they suffer from the problem of increased computational time as problem size increases. Training a QCNN requires repeated updates of a large number of variational parameters for a large number of training data sets, resulting in extremely long processing times. For example, a single parameter update per training data set requires O(n) (where n is the number of qubits). Furthermore, quantum computing is a developing technology, and solving large-scale problems with currently limited computing resources requires minimizing computational costs. Unlike classical CNNs, quantum computing requires multiple executions of a quantum circuit to calculate the average of the measurements to obtain a single output. This measurement inefficiency is one of the factors that reduces the computational efficiency of QCNNs.
[0019] Therefore, one of the objectives of this embodiment is to execute QCNN more efficiently. In this embodiment, for example, as shown in FIG. 1, in order to effectively utilize quantum bits that are discarded in the P layer of a conventional quantum circuit of QCNN, the circuit is branched, and subsequent calculations are executed in parallel even for quantum bits that would otherwise be discarded. This makes it possible to measure the quantum states of many quantum bits at once, improving the measurement efficiency of QCNN and, in turn, the execution efficiency. Here, quantum bits that are discarded in the P layer are, for example, quantum bits in the P layer of the quantum circuit of QCNN shown in FIG. 1 that are not connected to the next C layer and are therefore disconnected. Also, as shown in FIG. 1, the U of the FC layer 5 In addition, in order to realize parallel computation by branching, it is required that the same value is returned at all branch destinations when the circuit is branched, which requires the operation of translational symmetry for each layer of the QCNN.
[0020] FIG. 2 is a diagram for explaining the translational symmetry and branching of QCNN. FIG. 2 shows a quantum circuit of QCNN that is branched by translational symmetry in order to effectively utilize quantum bits that would otherwise be discarded. As shown in FIG. 2, U 1 ~U 3 The circuit is branched without discarding each quantum bit that has been subjected to a unitary transformation in step 1. In this case, each layer is made translationally symmetric so that the same value is returned at all branching points, and the same unitary transformation is applied to each branch after branching.
[0021] In this embodiment, preconditions are imposed on input data so that the same value is returned at all branch destinations. The overall processing flow of this embodiment is shown in Figure 3. Figure 3 is a flowchart showing the overall processing flow including checking of preconditions in this embodiment.
[0022] In Fig. 3, the processing flow of this embodiment begins with step S102, but before that, prerequisites are checked in step S101. In step S101, as a prerequisite check, for example, as shown in Fig. 3, it is checked whether the input quantum state is translationally symmetric, that is, whether the input quantum state can return the same value at all branch destinations when the circuit is branched. Note that whether translational symmetry is present can be determined, for example, using a Hamiltonian formula, which is an existing algorithm.
[0023] 4 is a diagram for explaining the preconditions of this embodiment. The precondition of this embodiment is that the input quantum state, i.e., the input data, is translationally symmetric, which means that, for example, when quantum bits are arranged periodically in one dimension as shown in FIG. 4, the following formula (3) holds:
[0024]
[0025] In the formula (3), for example, ρ in is the input quantum state, and T is the unitary transformation that translates one quantum bit. Using the formula (3), as shown in FIG. 4, even if the input quantum state is translated by one, ρ in and T † ρ in If T is equal to , the input quantum state is determined to be translationally symmetric.
[0026] If the input quantum state is translationally symmetric (step S101: Yes), proceed to this embodiment, where each layer of the QCNN is translationally symmetric (step S102) and branched (step S103). Then, training is performed using training data using the translationally symmetric branched QCNN (step S104). Furthermore, predictions for unknown data are made using the translationally symmetric branched QCNN trained in step S104 (step S105). After step S105 is executed, the overall processing shown in FIG. 3 ends.
[0027] On the other hand, if the input quantum state is not translationally symmetric (step S101: No), the process proceeds to the conventional method, and a QCNN as shown in Figure 1 is designed (step S106). Then, training is performed using training data using the QCNN as shown in Figure 1 (step S107). Furthermore, predictions for unknown data are made using the QCNN trained in step S107 (step S108). After step S108 is executed, the overall process shown in Figure 3 ends.
[0028] (Functional Configuration of Information Processing Device 10) Next, the functional configuration of the information processing device 10, which is the subject of operation in this embodiment, will be described. FIG. 5 is a diagram showing an example of the configuration of the information processing device 10 according to this embodiment. The information processing device 10 shown in FIG. 5 is, for example, a quantum computer, a server computer, or an information processing device such as a desktop PC (Personal Computer) or a notebook PC. Note that while FIG. 5 shows the information processing device 10 as a single computer, it may also be a distributed computing system consisting of multiple computers. Furthermore, the information processing device 10 may also be a cloud computing device managed by a service provider that provides cloud computing services.
[0029] As shown in FIG. 5, the information processing device 10 includes a communication unit 20, a storage unit 30, and a control unit 40.
[0030] The communication unit 20 is a processing unit that controls communication with other devices, and is, for example, a communication interface such as a network interface card or a USB (Universal Serial Bus) interface.
[0031] The storage unit 30 has a function of storing various data and programs executed by the control unit 40, and is realized by a storage device such as a memory or a hard disk. The storage unit 30 stores quantum circuit information 31, quantum state information 32, and the like.
[0032] For example, information about a quantum circuit of a QCNN is stored in the quantum circuit information 31. More specifically, for example, information about a quantum circuit of a QCNN branched in translational symmetry as shown in FIG.
[0033] Here, we will explain the translational symmetry of each layer of the QCNN. FIG. 6 is a diagram for explaining the translational symmetry of each layer of the QCNN according to this embodiment. As shown in FIG. 6, the quantum circuit of the QCNN is configured so that the convolutional / fully connected layer is symmetric with respect to the translational operation of one quantum bit. Specifically, the translational symmetry is performed by an appropriate combination of gate operations on the quantum bit. This point will be explained using a translationally symmetric QCNN layer.
[0034] FIG. 7 is a diagram showing an example of a translationally symmetric QCNN layer according to this embodiment. FIG. 7 also shows an example of a one-dimensional translationally symmetric layer. As shown in FIG. 7 , in each convolutional / fully connected layer, an X rotation gate, a Z rotation gate, and an X rotation gate are successively applied to all quantum bits, and a ZZ rotation gate is applied to adjacent quantum bits, and this operation is repeated d times. The rotation angles of the rotation gates applied in parallel are assumed to be the same. Translational symmetry of the layer is achieved by such an operation. The rotation angles when applying the X rotation gate and the Z rotation gate are variational parameters that must be updated repeatedly for a large amount of training data during QCNN training.
[0035] The branching of a quantum circuit of a QCNN branched by translational symmetry will be described. FIG. 8 is a diagram for explaining the branching of a QCNN according to this embodiment. The portion surrounded by a dashed line in FIG. 8 is the portion of the QCNN originally to be calculated. Therefore, in a conventional QCNN quantum circuit, quantum bits other than the portion of the QCNN originally to be calculated are discarded in the pooling layer. Also, only one of the quantum bits after operation in the fully connected layer is measured, and the other is discarded. Therefore, in this embodiment, as shown in FIG. 8, the quantum bits other than the portion of the QCNN originally to be calculated are not discarded, and the circuit is branched, and the set of quantum bits after branching is subjected to exactly the same operation as the portion of the QCNN originally to be calculated.
[0036] 8, for example, in a conventional QCNN quantum circuit, only one measurement value of the QCNN portion that is originally intended to be calculated can be obtained in one execution, whereas in this embodiment, eight measurement values of eight quantum bits can be obtained. This allows the information processing device 10 to execute QCNN more efficiently. Furthermore, although the amount of processing increases because quantum bits are used rather than discarded, and although it may take some time to rearrange the quantum bits in a quantum computer method such as superconductivity, all operations after branching can be performed in parallel, so there is almost no increase in calculation time due to branching.
[0037] Returning to the explanation of FIG. 5 , the quantum state information 32 stores, for example, information about the quantum state, which is the state of the quantum bit in the quantum computer that is the information processing device 10. More specifically, the quantum state information 32 stores, for example, information about the measured quantum state. Note that the information about the measured quantum state may include, in addition to the measurement value of the quantum state of the quantum bit in the final state, the average value of the measurement value to be used as the output of the QCNN, the gradient of the calculated output of the QCNN, and the like. The quantum state information 32 may also store, for example, information about training data.
[0038] The above information stored in the storage unit 30 is merely an example, and the storage unit 30 can store various information other than the above information.
[0039] The control unit 40 is a processing unit, such as a processor, that controls the entire information processing device 10. The control unit 40 includes a dividing unit 41, an executing unit 42, a calculating unit 43, and a setting unit 44. Each processing unit is an example of an electronic circuit included in the processor or an example of a process executed by the processor.
[0040] The division unit 41, for example, divides the multiple quantum bits operated in each layer of the QCNN into multiple quantum bit sets so that they are symmetric with respect to the translation operation. Here, the multiple quantum bits operated in each layer of the QCNN are, for example, multiple quantum bits input to a quantum circuit designed so that each layer of the QCNN is symmetric with respect to the quantum bit translation operation. Furthermore, the division of the multiple quantum bits by the division unit 41 is, for example, as shown in FIG. 8, dividing the multiple quantum bits so that the same operation is performed on each of the quantum bit sets after branching. While FIG. 8 shows an example of 8 quantum bits, division of n quantum bits is also possible.
[0041] FIG. 9 is a diagram for explaining the branching of n quantum bits according to this embodiment. FIG. 9 also shows an example of branching n quantum bits into d branch destinations (n is any natural number, and d is a divisor of n). The dividing unit 41, for example, divides n quantum bits into "n / d blocks of d quantum bits." The quantum bits of each block are branched into different branches of the quantum circuit of the QCNN, and after branching, the same operation is performed on all branches. The dividing unit 41 can also divide quantum bits in a similar manner, for example, in higher dimensions.
[0042] FIG. 10 is a diagram for explaining branching of multiple quantum bits in the case of a high dimension according to this embodiment. FIG. 10 illustrates branching of 2×3 blocks of a two-dimensional square lattice. For example, as shown by the dashed lines in the lower part of FIG. 10, the dividing unit 41 divides the two-dimensional square lattice of quantum bits into small blocks of 2×3, each containing six quantum bits. In this case, the division is performed so that the blocks are translationally symmetric. The six quantum bits in each block are branched into different branches, and after branching, the same operation is performed on all branches.
[0043] A description will now be given of the flow of branching processing of the QCNN, including the division of quantum bits by the division unit 41. Fig. 11 is a flowchart showing the flow of branching processing of the QCNN according to this embodiment.
[0044] 11, the information processing device 10 divides, for example, n quantum bits into mini-blocks each consisting of d quantum bits (step S201). At this time, the mini-blocks are set to be translationally symmetric with each other. Furthermore, d is set to be a divisor of n.
[0045] Next, the information processing device 10 branches, for example, the d quantum bits in each mini-block divided in step S201 into different branches of the quantum circuit of the QCNN (step S202).
[0046] Then, the information processing device 10 performs the division process of step S201 and the branching process of step S202 for each quantum bit after the operation in the branching destination layer until the number of quantum bits is sufficiently reduced (step S203). Here, "until the number of quantum bits is sufficiently reduced" means, for example, until the number of quantum bits to be divided and branched is two, as in the quantum circuit of QCNN shown in Figure 8. After executing step S203, the branching process shown in Figure 11 ends.
[0047] 5, the execution unit 42, for example, branches each of the multiple quantum bits in the quantum bit set divided by the division unit 41 into different branches of the quantum circuit to perform operations in each layer of the QCNN. Then, the execution unit 42, for example, measures the quantum states of the multiple quantum bits in the final state. Note that the division process by the division unit 41 and the execution process of the operation by the execution unit 42 are performed, for example, for each convolutional / fully connected layer of the quantum circuit of the QCNN.
[0048] The calculation unit 43, for example, calculates the average value of the measurement values of the quantum states measured by the execution unit 42. FIG. 12 is a diagram for explaining the calculation of the average of the measurement values of the quantum bits according to this embodiment. As shown in FIG. 12, the calculation unit 43, for example, calculates the average value of the measurement values of the quantum states of multiple quantum bits in the final state measured by the execution unit 42. As shown in FIG. 12, in this embodiment, multiple measurement values are obtained by one execution of the QCNN quantum circuit, and the average value of the multiple measurement values is calculated. Furthermore, the calculation unit 43, for example, inputs multiple quantum bits into the QCNN quantum circuit a predetermined number of times, and calculates the average value of the average values of the multiple measurement values calculated for each execution. That is, in this embodiment, the QCNN quantum circuit is executed a predetermined number of times, and the average of the average values of the measurement values calculated for each execution is calculated as the expected value of the measurement value.
[0049] As described above, in this embodiment, the average value of multiple measurements is used as the output of the QCNN, but the reason why the average value can be taken, that is, the equivalence of multiple measurements, will be explained. Figure 13 is a diagram (1) for explaining the equivalence of the measurements of quantum bits according to this embodiment. Figure 13 shows an example in which one quantum bit is translated by a translation operation. As shown in Figure 13, even if one quantum bit is translated, only the output location shifts, and the input state and circuit structure themselves remain unchanged, and the outputs of each quantum bit are all the same.
[0050] 13 shows an example of one quantum bit translation, but the same applies to the case of translating more than two quantum bits, such as two quantum bits. Fig. 14 is a diagram (2) for explaining the equivalence of the quantum bit measurement values according to this embodiment. As shown in Fig. 14, even when two quantum bits are translated by a translation operation, just as in the case of the one quantum bit translation shown in Fig. 13, only the output position shifts, and the input state and the circuit structure themselves remain unchanged, and the outputs of each quantum bit are all the same.
[0051] Therefore, in this embodiment, by executing the quantum circuit of the QCNN and measuring all the quantum bits output, it is possible to obtain more measurement values than a conventional quantum circuit, and the measurement efficiency of the QCNN and therefore the execution efficiency can be improved. Note that, regarding the measurement efficiency of this embodiment, even if n times more measurement values are obtained than a conventional quantum circuit, these measurement values are correlated with each other due to quantum entanglement, so the measurement efficiency does not necessarily become n times.
[0052] The flow of the output measurement process as described above will be described below. Fig. 15 is a flowchart showing the flow of the output measurement process by the translationally symmetric branched QCNN according to this embodiment.
[0053] 15, the information processing device 10 inputs, for example, a quantum state of n qubits to a quantum circuit of a QCNN branched in translational symmetry (step S301). Note that steps S301 to S303 are performed for the number of shots N, which is the number of executions of the specified quantum circuit. s Repeated for minutes.
[0054] Next, the information processing device 10 performs quantum measurements on each of the n quantum bits in the final state, for example, to obtain n measurement values (step S302).
[0055] Next, the information processing device 10 calculates the average of the n measured values obtained in step S302, for example, and designates it as the average value M i (Step S303). Then, the information processing apparatus 10 performs steps S301 to S303 for the number of shots N s Repeat for N s Average value M i get.
[0056] Next, the information processing device 10 performs, for example, s Average value M i The average value of (=Σ(M i / N s )) and set it as the output of the QCNN (step S304). After step S304 is executed, the measurement process shown in FIG. 15 ends.
[0057] Returning to the description of FIG. 5 , the setting unit 44 sets independent parameters for each of a plurality of rotation gates of the quantum circuit, which is controlled by one variational parameter of the loss function and is designed so that each layer of the QCNN is symmetric with respect to the translation operation of the quantum bits.
[0058] Furthermore, the calculation unit 43 fixes the parameters set by the setting unit 44 to one variational parameter of the loss function, for example. The calculation unit 43 also calculates the derivative of a predetermined number of measurement expectation values of a quantum circuit designed to make each layer of the QCNN symmetrical with respect to the translation operation of the quantum bits, with respect to one of the parameters. The calculation unit 43 also calculates the gradient of the output of the QCNN, for example, based on the predetermined number of differential values, the number of parameters, and the predetermined number. Note that the calculation of the gradient includes a process of calculating the gradient by, for example, calculating the formula "sum of the predetermined number of differential values × number of parameters / predetermined number." The gradient measurement method of the QCNN output of this embodiment will be described in more detail.
[0059] In the gradient measurement method of the QCNN output of this embodiment, a method of differentiating the QCNN output with a variation parameter α is used. FIG. 16 is a diagram for explaining the gradient measurement method of the QCNN output according to this embodiment. As will be described in detail later, in the method of differentiating the QCNN output with a variation parameter α, as shown in FIG. 1 Considering only the differential of α, the predetermined number n of measurements of the branched QCNN 1 The derivative at α is calculated by the parameter shift method or the like. 1 The equation "sum of n differential values (= each measurement value) × m / n" is calculated, and the calculation result is used as the gradient of the QCNN output. Note that, as will be described in detail later, m is the number of gates controlled by the variation parameter α, and in this embodiment, each gate is controlled by an independent parameter (α 1 ~α m), the number of parameters is m. In this embodiment, as shown in FIG. 16, even in the gradient measurement method, multiple measurement values can be obtained and used by executing the quantum circuit of QCNN once, so the measurement efficiency and therefore the execution efficiency of QCNN can be improved.
[0060] The gradient measurement method of the output of the QCNN in this embodiment will be described in more detail. In this embodiment, the gradient measurement is performed by the chain rule, and for example, as shown in the following equation (4), the loss function L({〈Z1(α)〉 i}) with the variational parameter α of QCNN to calculate the gradient.
[0061]
[0062] In equation (4), for example, i is the label of the training data, N t is the number of training data. i can be calculated as shown in equation (2), and 〈Z1〉 i =〈φ i |U † Z1U|φ1〉. In addition, in equation (4), ∂L / ∂〈Z1〉 i is the definition of L and 〈Z1〉 i Since it is determined only by the value of ∂〈Z1〉 i In this embodiment, the derivative of the loss function shown in equation (4) is calculated by quantum computation.
[0063] Here, in the translationally symmetric QCNN of this embodiment, as shown in FIG. 7, one parameter (e.g., α) controls multiple gates. Therefore, a gradient measurement method when one parameter controls multiple gates will be described. FIG. 17 is a diagram (1) for explaining a gradient measurement method when one parameter controls multiple gates according to this embodiment. As shown in FIG. 17, in this embodiment, for example, one parameter α that controls multiple gates is virtually a parameter α where each of the parameters of the multiple gates is a different parameter α. i (i=1 to m), where m is the number of gates controlled by the variation parameter α, as described above.
[0064] And α i Let be a function of α, and α i By setting (α) = α and using the chain rule of differentiation, we obtain ∂〈Z1〉 / ∂α as shown in the following equation (5).
[0065]
[0066] Note that ∂〈Z1〉 / ∂α i can be calculated by a conventional algorithm such as the parameter shifting method, as shown in the following equation (6).
[0067]
[0068] In equation (6), 〈Z1〉| αi=α±π / 4 For example, α i Only α±π / 4 is used, and other α j(≠i) is the expected value of Z1 in the quantum circuit with α.
[0069] However, to find ∂〈Z1〉 / ∂α, α1~α m It is required to calculate the derivatives of each of the parameters α1 to α2. FIG. 18 is a diagram (2) for explaining the gradient measurement method when one parameter controls multiple gates according to this embodiment. For example, the parameters α1 to α2 of each gate in the area enclosed by the frame in FIG. 18 are m In this embodiment, the calculation of the derivatives for each of α1 to α m Therefore, the measurement efficiency of QCNN and therefore its execution efficiency can be improved even in the gradient measurement method.
[0070] In this way, in this embodiment, a single quantum circuit of QCNN can obtain multiple differential values and calculate the gradient, and the reason for this will be explained below. Fig. 19 is a diagram (1) for explaining the efficiency measurement of the gradient using a branched QCNN according to this embodiment. Fig. 19 shows an example in which one quantum bit is translated by a translation operation and the differential position is moved to α1. As shown in Fig. 19, even if one quantum bit is translated, only the location of the output is shifted. For example, if the differential position is fixed at α1, each of the branched outputs will be α1. iThis corresponds to the derivative of , and the gradient can be measured with a single quantum circuit.
[0071] 19 shows an example of one quantum bit translation, but the same applies to the case of translating more than two quantum bits, such as two quantum bits. Figure 20 is a diagram (2) for explaining the efficiency measurement of the gradient using a branched QCNN according to this embodiment. As shown in Figure 20, even when two quantum bits are translated by a translation operation, the gradient can be measured with one quantum circuit, just like the one quantum bit translation shown in Figure 19, with only the output position being shifted.
[0072] Furthermore, branching can also improve measurement efficiency for the differentiation of the second and subsequent layers of the QCNN. FIG. 21 is a diagram (3) for explaining the efficient measurement of gradients using a branched QCNN according to this embodiment. In the example of FIG. 21, each of the top four outputs of the quantum bit corresponds to a term in the differential formula. Similarly, each of the bottom four outputs of the quantum bit corresponds to a term in the differential formula. Then, by performing measurements on both the upper and lower branches, 4 × 2 = 8 measured values can be obtained, which improves measurement efficiency compared to conventional quantum circuits using QCNN.
[0073] The flow of the output gradient measurement process as described above will be described below. Fig. 22 is a flowchart showing the flow of the output gradient measurement process of the QCNN according to this embodiment.
[0074] First, as shown in FIG. 22, the information processing device 10 assigns an independent parameter (α) to each of a plurality of rotary gates controlled by α. 1 ~α m ) is set (step S401).
[0075] Next, the information processing device 10, for example, m = α, and the expected value of n measurements in the branched QCNN is 1 The differential at is calculated by the parameter shift method (step S402). The flow of the differential measurement process by the parameter shift method in step S402 will be described. Fig. 23 is a flowchart showing the flow of the differential measurement process by the parameter shift method according to this embodiment.
[0076] As shown in FIG. 23, the information processing device 10, for example, 1 = α ± π / 4, and perform QCNN measurements. ± i,l (l=1 to n) (step S501). In step S501, the number of shots N s Repeated for minutes.
[0077] Next, the information processing device 10 calculates, for example, the average value of the differences in the measurement values (=Σ i (z + i,l -z - i,l ) / N s ) as the α of the lth measurement 1 After step S502 is executed, the differential measurement process shown in Fig. 23 ends, and the process returns to the gradient measurement process shown in Fig. 22.
[0078] Next, the information processing device 10, for example, 1 Then, the gradient measurement process shown in FIG. 22 is completed (step S403).
[0079] 5, the information processing device 10 trains, for example, a translationally symmetric branched QCNN using training data that is a set of quantum states of multiple quantum bits and a predetermined ground truth label. Furthermore, the information processing device 10 inputs the quantum states of multiple quantum bits to the translationally symmetric branched QCNN trained using the training data, and makes predictions for unknown data.
[0080] This embodiment, which allows QCNN to be executed more efficiently, has been described above. Below, we present numerical verification results to demonstrate the effects of this embodiment. In particular, we numerically demonstrate for a specific problem that the QCNN has sufficient performance to solve classification tasks even when translational symmetry is imposed on the QCNN. Furthermore, we demonstrate the improvement rate of measurement efficiency due to branching of the QCNN for this specific problem. Furthermore, the specific problem is set as a quantum phase classification. More specifically, the QCNN classifies whether the ground state of the Hamiltonian belongs to the SPT (Symmetry Protected Topological) phase (a symmetry-protected topological phase). The Hamiltonian can be expressed by the following equation (7):
[0081]
[0082] In formula (7), X j and Z j are, for example, the Pauli X and Z matrices for the jth qubit, h1 and h2 are real numbers, and L is a positive integer. Here, the periodic boundary condition (X j =X j+L , Z j =Z j+L ) Since the Hamiltonian expressed by equation (7) satisfies translational symmetry, (T † Since the ground state, which is the input data, is translationally symmetric, it satisfies the conditions for applying this embodiment (the prerequisites for this embodiment described with reference to FIG. 3). It is known that the phase diagram of the ground state of a Hamiltonian is, for example, as shown in FIG. 24. FIG. 24 is a diagram showing an example of the phase diagram of the ground state of a Hamiltonian.
[0083] The circuit structure is a translationally symmetric branched QCNN quantum circuit as shown in Figure 2, and each U i is created as shown in FIG. 7, and the simulation is performed assuming that there are 4×5=20 variation parameters with d=5.
[0084] The loss function is expressed by the following equation (8).
[0085]
[0086] In equation (8), U(θ) is, for example, the entire unitary transformation of QCNN, and Z1 is the physical quantity to be measured. i 〉 is N, as shown in FIG. 25, from h1=0.05 to 1.95 on the h2=0 axis. t = 20. Figure 25 shows an example of the relationship between the phase diagram of the ground state of the Hamiltonian and the training data of the QCNN. i If h1<1, then y i =1 (SPT phase), if h1>1, y i = 0 (except for the SPT phase). It is known that on the h2 = 0 axis, a phase transition occurs at h1 = 1.
[0087] The simulation conditions are as follows: L = 8, 12, and 16 qubit systems are simulated on a classical computer, 50 initial values for the variational parameters are randomly selected, and their average is calculated. The optimization method uses the existing Adam algorithm, and the simulation is performed with an infinite number of shots. However, measurement efficiency is estimated with a finite number of shots. The method for estimating measurement efficiency will be described later. It is generally believed that translational symmetry of the circuit reduces the expressive power of QCNN. As a solution to this problem, the correct classification label is set to a value between ±1, which is the spectral edge of the Pauli operator, rather than to a value between ±1 (e.g., 0.5).
[0088] Next, we will show performance evaluation using the above-mentioned simulation. Figure 26 is a diagram showing an example of the change in the loss function due to variational parameter update. As shown in Figure 26, the loss function decreases through optimization and ultimately converges to a small value as the problem size L increases. This suggests that the translationally symmetric branched QCNN according to this embodiment has the training potential to be applied to large-scale problems.
[0089] Figure 27 shows examples of prediction results for various problem sizes using a trained QCNN for training data. Figure 27 shows prediction results for training data on the h2 = 0 axis, i.e., data already known to the QCNN, using a trained QCNN. In Figure 27, the dashed line indicates the correct label; closer to the dashed line, the better the training. For quantification, we consider 0.5, exactly midway between the correct label y = 0 and 1, as the reference for the phase transition. For L = 16, <Z1> = 0.5 for h1 to 0.95, which means that the phase transition is correctly detected with an error of approximately 5% compared to the true phase transition point, h1 = 1. This result demonstrates that the translationally symmetric branched QCNN of this embodiment also has sufficient expressive power to solve the task of using a QCNN to classify whether the ground state of a Hamiltonian belongs to the SPT phase.
[0090] FIG. 28 shows an example of prediction results for unknown data for each problem size using a trained QCNN. FIG. 28 shows prediction results for a phase diagram on the h1-h2 plane using a trained QCNN. While FIG. 27 shows prediction results for data on the h2=0 axis, which is known to the QCNN, FIG. 28 shows prediction results for data on a non-h2=0 axis, which is unknown to the QCNN. According to the prediction results in FIG. 28, the lower phase transition (SPT phase → antiferromagnetic phase), which was not included in the training data, can also be detected. This indicates that the translationally symmetric branched QCNN according to this embodiment also has sufficient generalization capabilities.
[0091] The above results show that the translationally symmetric branched QCNN according to this embodiment has sufficient trainability, expressive power, and generalization ability to solve quantum phase classification tasks.
[0092] Note that the expected value estimated from an experiment with a finite number of shots generally differs from the true expected value due to statistical error. When estimating the expected value from the average of the measured values, the probability distribution of the estimated value is (average = true expected value, standard deviation ∝ 1 / √N s) is a normal distribution. Such a probability distribution of the estimated value is shown in FIG. 29. FIG. 29 is a diagram showing an example of the probability distribution of the estimated value. As shown in FIG. 29, the ratio of the measurement efficiency of the branched QCNN to that of the non-branched QCNN is, for example, r=(σ0 / σ t ) 2 are defined as follows: σ0 and σ t and denote the standard deviation of the expected value estimates for the unbranched and branched QCNNs, respectively. Also, r = (σ0 / σ t ) 2 The meaning of the index is that when you want to estimate the expected value with a certain accuracy, the number of shots required for a branched QCNN is 1 / r times larger than that for a non-branched QCNN. t is estimated by, for example, performing a simulation of "estimating <Z1> from the results of 1000 shots" 10,000 times. Note that a branched QCNN and a non-branched QCNN are shown in Figure 30. As shown in Figure 30, in the quantum circuit of a non-branched QCNN, there is one measurement point for the quantum state of the qubit, whereas in the quantum circuit of a branched QCNN, there are n points (8 points in the example of Figure 30). Also, in the branched QCNN and the non-branched QCNN, i is common.
[0093] Figure 31 shows an example of the measurement efficiency due to branching of QCNN for three representative states. As shown in Figure 31, the measurement efficiency of the SPT state decreases with training, eventually reaching about 2.5 to 3 times, while no significant decrease in measurement efficiency can be observed for the other two states.
[0094] Figure 32 shows another example of the measurement efficiency due to branching of the QCNN for three representative states. Figure 32 plots the measurement efficiency for three representative states after training as the improvement rate of measurement efficiency due to branching of the QCNN. Linear fitting is also performed for each of SPT, PM (paramagnetic), and AFM (antiferromagnetic). As a result, as shown in Figure 32, linear fitting is possible within the range of the error bars, which means that the measurement efficiency is improved by, for example, O(n) times (n is the number of quantum bits) due to branching of the circuit.
[0095] As explained above, in a QCNN, by performing translational symmetry and branching on the quantum circuit as in this embodiment, the measurement efficiency can be increased by O(n) times (n is the number of quantum bits). Furthermore, verification results for the translationally symmetric branched QCNN according to this embodiment showed that it is possible to detect quantum phase transitions, and it was estimated that the measurement efficiency would be increased by O(n) times.
[0096] (Effect) As described above, the information processing device 10 inputs the quantum states of each of the multiple quantum bits into a quantum circuit designed so that each layer of the QCNN is symmetric with respect to a translation operation of the quantum bits, performs an operation in the first layer of the QCNN on the multiple quantum bits, divides the multiple quantum bits operated in the first layer into multiple quantum bit sets so that the multiple quantum bits are symmetric with respect to the translation operation, branches each of the multiple quantum bits in the quantum bit set into a different branch of the quantum circuit to perform an operation in each layer of the QCNN, and measures the quantum state of each of the multiple quantum bits in the final state.
[0097] In this way, the information processing device 10 divides multiple quantum bits input to the translationally symmetric quantum circuit of the QCNN into quantum bit sets, branches each quantum bit into a different branch, performs operations on each layer, and measures the quantum state. This allows the information processing device 10 to more efficiently execute the QCNN.
[0098] Furthermore, the information processing device 10 calculates the average value of the measurement values of the quantum states.
[0099] This allows the information processing device 10 to execute QCNN more efficiently.
[0100] Furthermore, the information processing device 10 inputs a plurality of quantum bits into the quantum circuit a predetermined number of times and calculates the average value of the average values calculated for each input.
[0101] This allows the information processing device 10 to execute QCNN more efficiently.
[0102] In addition, the information processing device 10 sets independent parameters for each of a plurality of rotation gates of the quantum circuit, which are controlled by one variational parameter of the loss function, fixes the parameters to the variational parameters, calculates the derivative of one of the parameters of a predetermined number of measurement expectation values of the quantum circuit, and calculates the gradient of the output of the QCNN based on the predetermined number of differential values, the number of parameters, and the predetermined number.
[0103] This allows the information processing device 10 to execute QCNN more efficiently.
[0104] Furthermore, the process of calculating the gradient executed by the information processing device 10 includes a process of calculating the gradient by calculating the sum of a predetermined number of differential values x the number of parameters / a predetermined number of equations.
[0105] This allows the information processing device 10 to execute QCNN more efficiently.
[0106] Furthermore, the information processing device 10 trains the QCNN using training data that is a pair of quantum states of multiple quantum bits and a predetermined correct answer label.
[0107] This allows the information processing device 10 to execute QCNN more efficiently.
[0108] Furthermore, the information processing device 10 inputs the quantum states of multiple quantum bits into a QCNN trained using training data, and makes predictions for unknown data.
[0109] This allows the information processing device 10 to execute QCNN more efficiently.
[0110] (System) The information, including the processing procedures, control procedures, specific names, various data, and parameters shown in the above documents and drawings, may be changed as desired unless otherwise specified. Furthermore, the specific examples, distributions, and numerical values described in the embodiments are merely examples and may be changed as desired.
[0111] Furthermore, the specific form of distribution or integration of the components of the information processing device 10 is not limited to that shown in the figure. For example, the execution unit 42 of the information processing device 10 may be distributed across multiple processing units, or the execution unit 42 and the calculation unit 43 of the information processing device 10 may be integrated into a single processing unit. That is, all or some of the components may be functionally or physically distributed or integrated in any unit depending on various loads, usage conditions, etc. Furthermore, all or any part of the processing functions of each device may be realized by a CPU (Central Processing Unit) and a program analyzed and executed by the CPU, or may be realized as hardware using wired logic.
[0112] Fig. 33 is a diagram showing an example of the hardware configuration of an information processing device 10 according to this embodiment. As shown in Fig. 33, the information processing device 10 includes a communication interface 10a, a hard disk drive (HDD) 10b, a memory 10c, and a processor 10d. The components shown in Fig. 33 are connected to each other via a bus or the like.
[0113] The communication interface 10a is a network interface card or the like, and communicates with other information processing devices. The HDD 10b stores, for example, programs and data that operate the various functions of the information processing device 10.
[0114] The processor 10d is a CPU, a microprocessing unit (MPU), a graphics processing unit (GPU), or the like. The processor 10d may also be implemented by an integrated circuit such as an application specific integrated circuit (ASIC) or a field programmable gate array (FPGA). The processor 10d reads, for example, from the HDD 10b, a program that executes the same processes as those of the processing units shown in FIG. 5, and loads the program into the memory 10c. This allows the processor 10d to operate as a hardware circuit that executes processes that implement the functions of the information processing device 10.
[0115] Furthermore, the information processing device 10 can also realize the same functions as those of the above embodiment by reading the program from a recording medium using a medium reading device and executing the read program. Note that the program in this other embodiment is not limited to being executed by the information processing device 10. For example, the above embodiment may also be applied in the same way to cases where another information processing device executes the program or where the other information processing device and the information processing device 10 cooperate to execute the program.
[0116] The program may be distributed via a network such as the Internet. The program may also be recorded on a computer-readable storage medium such as a hard disk, a flexible disk (FD), a CD-ROM, a magneto-optical disk (MO), or a digital versatile disk (DVD). The program may then be read from the storage medium and executed by the information processing device 10 or the like.
[0117] REFERENCE SIGNS LIST 10 Information processing device 10a Communication interface 10b HDD 10c Memory 10d Processor 20 Communication unit 30 Storage unit 31 Quantum circuit information 32 Quantum state information 40 Control unit 41 Division unit 42 Execution unit 43 Calculation unit 44 Setting unit
Claims
1. An information processing program that causes a computer to execute the following processes: inputting the quantum state of each of a plurality of quantum bits into a quantum circuit designed so that each layer of a quantum convolutional neural network is symmetric with respect to a translation operation of the quantum bits; executing an operation in a first layer of the quantum convolutional neural network for the plurality of quantum bits; dividing the plurality of quantum bits operated in the first layer into a plurality of quantum bit sets so that the operation is symmetric with respect to the translation operation; branching each of the plurality of quantum bits in the quantum bit set into a different branch of the quantum circuit to execute an operation in each layer of the quantum convolutional neural network; and measuring the quantum state of each of the plurality of quantum bits in the final state.
2. The information processing program according to claim 1, characterized in that the computer is caused to execute a process of calculating an average value of the measurement values of the quantum state.
3. The information processing program according to claim 2, characterized in that the computer is caused to execute a process of inputting the plurality of quantum bits to the quantum circuit a predetermined number of times and calculating the average value of the average values calculated for each input.
4. The information processing program according to claim 1, characterized in that the computer is caused to execute the following processes: setting an independent parameter for each of a plurality of rotation gates of the quantum circuit, which are controlled by one variational parameter of a loss function; fixing the parameter to the variational parameter, and calculating the derivative of a predetermined number of measurement expectation values of the quantum circuit with respect to one of the parameters; and calculating the gradient of the output of the quantum convolutional neural network based on the predetermined number of derivative values, the number of parameters, and the predetermined number.
5. The information processing program according to claim 4, characterized in that the process of calculating the gradient includes a process of calculating the gradient by calculating the sum of the predetermined number of differential values x the number of the parameters / the predetermined number of equations.
6. An information processing program according to any one of claims 1 to 5, characterized in that the computer is caused to execute a process of training the quantum convolutional neural network using training data which is a pair of quantum states of the multiple quantum bits and a predetermined correct answer label.
7. The information processing program according to claim 6, characterized in that the quantum states of the multiple quantum bits are input to the quantum convolutional neural network trained using the training data to make predictions for unknown data.
8. An information processing device comprising a control unit that executes the following processes: inputting the quantum state of each of a plurality of quantum bits into a quantum circuit designed so that each layer of a quantum convolutional neural network is symmetric with respect to a translation operation of the quantum bits; executing an operation in a first layer of the quantum convolutional neural network for the plurality of quantum bits; dividing the plurality of quantum bits operated in the first layer into a plurality of quantum bit sets so that the operation is symmetric with respect to the translation operation; branching each of the plurality of quantum bits in the quantum bit set into a different branch of the quantum circuit to execute an operation in each layer of the quantum convolutional neural network; and measuring the quantum state of each of the plurality of quantum bits in a final state.
9. An information processing method characterized in that a computer executes the following processes: inputting the quantum state of each of a plurality of quantum bits into a quantum circuit designed so that each layer of a quantum convolutional neural network is symmetric with respect to a translation operation of the quantum bits; executing an operation in a first layer of the quantum convolutional neural network for the plurality of quantum bits; dividing the plurality of quantum bits operated in the first layer into a plurality of quantum bit sets so that the operation is symmetric with respect to the translation operation; branching each of the plurality of quantum bits in the quantum bit set into a different branch of the quantum circuit to execute an operation in each layer of the quantum convolutional neural network; and measuring the quantum state of each of the plurality of quantum bits in the final state.
Citation Information
Patent Citations
System and method for quantum processing of data
JP2019096334A
Differential device
JP7364785B2
System and method for providing data computation via quantum computers
US20220101164A1
Quantum processing of probabilistic numeric convolutional neural networks
US20220253741A1
Quantum convolutional neural networks
US20210383189A1
