Neural network and machine learning method

The neural network with a pseudo-quantum spin layer and error propagation from two previous layers addresses the challenge of achieving high efficiency and accuracy in backpropagation on general digital computers, effectively mimicking quantum mechanical effects for complex problems.

WO2024236704A1PCT designated stage expired Publication Date: 2024-11-21NT T INC
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
PCT/JP2023/018157
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Filing Date
2023-05-15
Publication Date
2024-11-21

AI Technical Summary

Technical Problem

Current backpropagation methods for neural networks struggle to achieve high efficiency and accuracy on general digital computers, especially when handling complex problems, due to the difficulty in reproducing quantum mechanical effects used in quantum annealing.

Method used

A neural network architecture incorporating a pseudo-quantum spin layer with a time evolution equation that approximates quantum annealing, allowing error propagation from two previous layers, enabling high computational efficiency and accuracy on general digital computers.

Benefits of technology

The proposed method achieves high accuracy in a short time, similar to methods directly introducing quantum mechanical effects, while requiring less computational load, making it feasible on standard digital computers.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure JP2023018157_21112024_PF_FP_ABST
    Figure JP2023018157_21112024_PF_FP_ABST
Patent Text Reader

Abstract

The present invention proposes a neural network and a machine learning method that can implemented in general digital computers with efficiency comparable to that of a backpropagation method incorporating quantum mechanical effects. A neural network according to the present disclosure provides a neural network comprising, in the intermediate layers thereof, a spinning layer which contains an integer number of neurons that is a multiple of three and which generates output having the form of a spin operator, and a pseudo-quantum spin layer which is coupled to the spinning layer and which contains a plurality of layers internally, wherein the pseudo-quantum spin layer is configured to output spins on the basis of time-evolution-aware operations similar to quantum annealing.
Need to check novelty before this filing date? Find Prior Art

Description

Neural Networks and Machine Learning Methods

[0001] The present disclosure relates to neural networks and machine learning methods, and more particularly to backpropagation methods using neural networks.

[0002] Quantum annealing is a computational technique that applies quantum mechanical effects (such as tunneling effects) to perform combinatorial optimization. Numerical calculations that apply quantum mechanical effects have been attracting attention in the field of quantum computing in recent years due to their high versatility, and vigorous research has been conducted on them. However, it is considered difficult to reproduce the quantum mechanical effects used in quantum annealing using general digital computers.

[0003] Meanwhile, the backpropagation method (back error propagation method) is known as a typical learning method for neural networks that is often used in machine learning (see, for example, Non-Patent Document 1). The backpropagation method has become widely used as a basic learning method for deep learning, but in recent years, there has been an increasing demand for handling more complex problems, and therefore, further improvement in efficiency (i.e., achieving high accuracy in a short time) is required.

[0004] One existing technology that satisfies these demands is a method of introducing the above-mentioned quantum mechanical effects into neural networks (see, for example, Non-Patent Document 2). However, this learning method requires a dedicated quantum computer to implement, and it is difficult to achieve large-scale calculations. Therefore, there is a need for a backpropagation learning method that can be implemented on a general digital computer and can achieve high efficiency.

[0005] David E. Rumelhart, Geoffrey E. Hinton, and Ronald J. Williams, “Learning representations by back-propagating errors”, Nature 323 (6088), 533-536 (1986)Mohammad H. Amin, Evgeny Andriyash, Jason Rolfe, Bohdan Kulchytskyy, and Roger Melko, “Quantum Boltzmann Machine”, Phys. Rev. 8, 021050 (2018)T. Kadowaki and H. Nishimori, “Quantum annealing in the transverse Ising model” Phys. Rev. E 58, 5355 (1998)Aditi Misra-Speldenner, Dmitry Bagrets, Frank Wilhelm-Mauch, Peter Schuhmacher, Tim Bode, Tobias Stollenwerk “Semiclassical (mean field) Approximate Optimization Algorithm” Adiabatic quantum computing (2022)〈URL: http: / / video.ictp.it / WEB / 2022 / 2022_06_20-smr3718 / 2022_06_20-11_50-smr3718.mp4〉

[0006] The present disclosure has been made in consideration of the above-mentioned problems, and its purpose is to propose a neural network and a machine learning method that can be implemented on a general digital computer with efficiency equivalent to that of the backpropagation method that incorporates quantum mechanical effects.

[0007] In response to the above-described problems, the present disclosure provides a neural network that includes a spin layer that includes an intermediate layer including neurons whose number is an integer multiple of three and that generates an output having the form of a spin operator, and a pseudo-quantum spin layer that is coupled to the spin layer and includes a plurality of layers therein, wherein the spin layer generates an output having the form of the spin operator based on (Equation 16) and (Equation 17) described below, and the pseudo-quantum spin layer takes time evolution into account based on (Equation 19) described below, and is configured to output spins based on (Equation 18) and (Equation 20).

[0008] The present disclosure further provides a machine learning method using the above-described neural network, in which an error function is given by the following (Equation 28), and a weight matrix and a phase are updated based on propagation of errors from the previous layer and the previous two layers. In one embodiment, the machine learning method is a backpropagation method.

[0009] 1 is a diagram showing the conceptual structure of a neural network 100 according to the present disclosure. It is a diagram showing operations performed in a general layer and each layer constituting a neural network 100 according to the present disclosure, where (a) shows the operation in the general layer, (b) shows the operation in the spin layer 102, (c) shows the operation in the pseudo quantum spin layer 103, and (d) shows the operation in the output layer 104. It is a diagram conceptually showing the operation flow of a machine learning method using a conventional neural network, where (a) shows the procedure using a feedforward neural network and (b) shows the procedure using a backpropagation method. It is a diagram conceptually showing the operation flow of a machine learning method using a neural network 100 according to the present disclosure, where (a) shows the procedure using a feedforward neural network and (b) shows the procedure using a backpropagation method. It is a diagram conceptually showing the structure of a neural network 500 according to the present disclosure that is specialized for solving combinatorial problems. 1 is a diagram showing the accuracy rate in the MNIST benchmark using the backpropagation method according to the present disclosure and the backpropagation method according to the present disclosure. 2 is a diagram showing the results of predicting time-series data in the NARMA10 benchmark using the backpropagation method according to the present disclosure, where (a) shows the target data and predicted data, and (b) shows the error. 3 is a block diagram showing an example of the configuration of a computer 800 that implements the machine learning method according to the present disclosure using a neural network according to the present disclosure.

[0010] Various embodiments of the present disclosure will be described in detail below with reference to the drawings. The same or similar reference numerals indicate the same or similar elements, and redundant description may be omitted. Numerical values ​​are for illustrative purposes and are not intended to limit the technical scope of the present disclosure. Furthermore, the following description is an example, and some configurations may be omitted or modified, or additional configurations may be added, as long as they do not deviate from the gist of one embodiment of the present disclosure.

[0011] The neural network disclosed herein is characterized by incorporating quantum mechanical effects in a pseudo manner. Furthermore, the backpropagation method using such a neural network is extended by using a time evolution equation that approximately represents the quantum annealing method. This extended backpropagation method achieves high computational efficiency, similar to the backpropagation method of the prior art, which directly incorporates quantum mechanical effects.

[0012] Furthermore, the backpropagation method according to the present disclosure using the neural network according to the present disclosure has the characteristic that error propagation occurs from the layer two layers before. By using such a method, it is possible to achieve high accuracy in a short time even in machine learning (backpropagation method) in a neural network that incorporates the above-mentioned pseudo-quantum mechanical effects. In addition, because the backpropagation method according to the present disclosure incorporates "pseudo" quantum mechanical effects, the load on the computer is relatively low compared to methods that directly incorporate quantum mechanical effects, such as those in conventional technology. Therefore, it can be implemented on a general digital computer.

[0013] Before describing the details of the invention, a method for approximately describing quantum mechanical time evolution based on the principles of quantum annealing will be described in detail below. This method is already known, but the neural network and backpropagation method according to the present disclosure are further expanded based on the concept of this existing method.

[0014] (Existing Method for Approximately Describing Quantum Mechanical Time Evolution) Generally, in the quantum annealing method, the time evolution of the Hamiltonian H of the Ising model as shown in (Equation 1) is considered (see, for example, Non-Patent Document 3).

[0015]

[0016] where σ αi (α=x, y, z), σ αj(α=x, y, z) are the Pauli matrices of the x, y, and z components at the i-th (i=1,...,N) and j-th (j=1,...,N, where i≠j) coordinates, respectively, and are 2x2 matrices as expressed in (Equation 2). ij is a term that represents the interaction between the i-th (i = 1, ..., N) and j-th (j = 1, ..., N) spins, and h α (α=x, y, z) is a term representing the local magnetic field of each of the x, y, and z components.

[0017]

[0018] When the mean field approximation is performed on the Hamiltonian H of such an Ising model, (Equation 1) can be rewritten as (Equation 3).

[0019]

[0020] Here, 〈σ zi 〉 is σ zi (The calculation method will be described later.) Furthermore, part of the first term is zi =Σ i J ij <σ zi 〉, and to extend it to a neural network, h as the y component yi Adding the terms, we get (Equation 4).

[0021]

[0022] Each spin operator σ αi The small time evolution of the expectation value of (α=x, y, z) can be described by (Equation 5).

[0023]

[0024] Here, t is time, δt is an infinitesimal time, and |φ(t)〉 is the state at time t. Here, the time evolution operator described by an exponential function is expressed by (Equation 6).

[0025]

[0026] Furthermore, here, Δ i and f → iare given by (Equation 7) and (Equation 8), respectively.

[0027]

[0028] Furthermore, by using the vector notation of the Pauli matrix, the time evolution can be described by (Equation 9).

[0029]

[0030] Here, Ci, Si, |f in (Equation 9) i 〉〈f i | and Ri are given by (Equation 10) to (Equation 12), respectively.

[0031]

[0032] In this way, quantum mechanical time evolution can be approximately described (see, for example, Non-Patent Document 4).

[0033] Based on the above, the neural network and the backpropagation method using the same according to the present disclosure will be described in detail below.

[0034] (Structure of Neural Network) Fig. 1 is a diagram showing the conceptual structure of a neural network 100 according to the present disclosure. As shown in Fig. 1, the neural network 100 includes an input layer 101, an output layer 104, a spin layer 102 coupled to the input layer 101, and a pseudo quantum spin layer 103 coupled to the spin layer 102 and the output layer 104. The input layer 101 and the output layer 104 each have an arbitrary integer N input and N output The pseudo quantum spin layer 103 has a structure similar to that of the input layer and output layer of a general neural network, each containing neurons, and has an arbitrary activation function g(x). On the other hand, the spin layer 102 and the pseudo quantum spin layer 103 are layers that provide a pseudo quantum mechanical effect. Furthermore, as shown in FIG. 1, the pseudo quantum spin layer 103 further includes multiple layers therein.

[0035] 1, the neural network 100 is depicted as being composed of four layers, but any number of general neural network layers (hereinafter referred to as general layers) may be included between the input layer 101 and the spin layer 102, and between the pseudo quantum spin layer 103 and the output layer 104. In other words, the spin layer 102 and the pseudo quantum spin layer 103 may be inserted between any intermediate layers in a normal deep neural network.

[0036] 2A and 2B are diagrams showing operations performed in the general layer and each layer constituting the neural network 100 according to the present disclosure, where (a) shows the operation in the general layer, (b) shows the operation in the spin layer 102, (c) shows the operation in the pseudo quantum spin layer 103, and (d) shows the operation in the output layer 104. As shown in FIG. 2A, in the operation in the general layer, in an arbitrary layer (here, the l-th layer), an input x from the l-1-th layer is calculated. i (l-1) , as shown in (Equation 13), the weight parameter W ij (l) are multiplied, and then converted into a format to be input to the next layer using the activation function (Equation 14) and output.

[0037]

[0038] This is an operation equivalent to the linear transformation process for weighting and outputting to the next neuron, which is performed in each layer of a general neural network. ij (l) is a parameter determined by the backpropagation learning method.

[0039] On the other hand, the spin layer 102 in the neural network 100 according to the present disclosure is an integer multiple of 3, 3N spin pieces (N spin This means that the spin layer 102 includes N spin neurons expressed by (Equation 15). spin The weighting formula in the spin layer 102 is x, which is the output (neuron information) of the input layer 101. j(in) Using the above, it is expressed as (Equation 16).

[0040]

[0041] Here, W αij (in) is 3N spin ×N input The matrix (weight matrix) is determined by the backpropagation method. On the other hand, the activation function in the spin layer 102 is expressed by (Equation 17).

[0042]

[0043] In this way, the spin layer 102 converts the output of the general layer into the form of a spin operator using Equations 16 and 17. The converted output of the spin layer 102 is input to the next layer (pseudo quantum spin layer 103).

[0044] The pseudo quantum spin layer 103 is a layer in which the same time evolution as that of the mean-field approximation quantum annealing method described above occurs. However, unlike the general quantum annealing method, in the l-th layer of the pseudo quantum spin layer, the output σ αi (l-1) is used as an input, and weighting expressed by (Equation 18) is performed. The l=0th layer corresponds to the spin layer 102 described above.

[0045]

[0046] Here, W αiγi (l-1) is 3N spin ×3N spin is a weight matrix, which is determined by the backpropagation method. That is, (Equation 18) corresponds to the calculation of weighting in an arbitrary layer (the l-1th layer) of the pseudo quantum spin layer 103. Then, h obtained by (Equation 18) is αi (l) An operation (transformation) is performed on the activation function, and the transformed output (spin) is input to the next layer (lth layer).

[0047] The spin input to this next layer (the output of the previous layer) is described by (Equation 19), which is a pseudo-expression of quantum annealing, similar to (Equation 9) above.

[0048]

[0049] Here, 2δtΔ in (Equation 10) i The term is replaced by a new variable φ i (l) Then, (Equation 20) is obtained. Note that this new variable φ i (l) is also determined by the backpropagation method. Furthermore, from the above (Equation 7) and (Equation 8), the relationship of (Equation 20) holds.

[0050]

[0051] Therefore, by using (Equation 20), |f in (Equation 19) i (l) 〉〈f i (l) | and R i (l) is given in the same matrix notation as (Equation 11) and (Equation 12). This (Equation 20) plays the role of an activation function in each layer of the pseudo quantum spin layer 103.

[0052] Here, the evolution of the spin is as follows: f i (l) , φ i (l) This can be summarized as a time evolution matrix described as a function of

[0053]

[0054] The spins output from the final layer (the layer connected to the output layer 104, here assumed to be the Lth layer, for example) in the pseudo quantum spin layer 103 become the input to the output layer 104. Then, in the output layer 104, weighting is performed according to (Equation 22), and the final learning output is generated by calculation using the activation function shown in (Equation 23). Note that the activation function here may be a soft matrix function or the like.

[0055]

[0056] Thus, the neural network 100 according to the present disclosure includes a spin layer 102 that generates an output converted into the form of a spin operator, and a pseudo quantum spin layer 103 that receives the output of the spin layer 102 as an input and generates a time evolution similar to that of a mean-field approximation quantum annealing method. Because the operations in the spin layer 102 and the pseudo quantum spin layer 103 provide a pseudo quantum mechanical effect, it becomes possible to implement a general digital computer with high computational efficiency similar to that of a machine learning method according to conventional technology that directly introduces quantum mechanical effects.

[0057] (Machine Learning Method) Next, a machine learning method using the neural network 100 according to the present disclosure will be described in detail below. For comparison, a conventional machine learning method will be described first.

[0058] 3 is a diagram conceptually showing the flow of calculations in a conventional machine learning method using a neural network, where (a) shows the procedure using a feedforward neural network and (b) shows the procedure using the backpropagation method. In a feedforward (forward propagation) neural network, the output from one layer is input to the next layer, and a linear transformation is performed using the weighting given by (Equation 13) and the activation function given by (Equation 14). This procedure is repeated for any number of layers, and the final output x i out The weight parameter W of this neural network is obtained. ij (l) is updated to make the final output the desired one. Therefore, a specific loss function E is considered using the final output, and calculations are performed to minimize this. For example, when using training data for image recognition, cross entropy or the like is set as the loss function E. Then, the weight parameters are updated according to (Equation 24) using the steepest descent method.

[0059]

[0060] The backpropagation method is a method of performing calculations for (Equation 24) taking (Equation 25) into consideration.

[0061]

[0062] Here, δ i (l) is called the error function and is given by (Equation 26).

[0063]

[0064] The backpropagation method is characterized by the fact that this error function is transmitted sequentially from the later layer to the earlier layer. The error function of the final output is determined by the loss function and the activation function of the final layer, and is given by (Equation 27) in the case of a combination of cross entropy and softmax function, for example.

[0065]

[0066] where t k is the training data.

[0067] Next, a machine learning method according to the present disclosure using the neural network 100 according to the present disclosure will be described.

[0068] 4A and 4B are diagrams conceptually illustrating the computational flow of a machine learning method using a neural network 100 according to the present disclosure, with (a) illustrating the procedure using a feedforward neural network and (b) illustrating the procedure using the backpropagation method. As shown in FIG. 4A, the input / output flow itself is similar to that of a machine learning method using a general neural network. On the other hand, the backpropagation method differs from conventional techniques in that error propagation also occurs from the layer two layers prior, as shown in FIG. 4B.

[0069] In the backpropagation method according to the present disclosure, the error function is given by (Equation 28).

[0070]

[0071] Here, the relationship between (Equation 29) and (Equation 33) holds.

[0072]

[0073] Furthermore, ε αβγ is the symbol for Levi-Civita.

[0074] The second term in (Equation 28) is a term that indicates the propagation of errors from the layer two layers before. By performing such error propagation, it becomes possible to achieve high accuracy in a short time.

[0075] Furthermore, in the backpropagation method according to the present disclosure, the sum of γ (= x, y, z) occurs within the same layer of the neural network used, so correlation occurs within the same layer, and learning is performed in the same way as when a recurrent neural network is used.

[0076] After the error to be propagated is obtained by (Equation 28), the weight matrix is ​​updated by (Equation 34) in the same way as (Equation 24) in the conventional backpropagation method.

[0077]

[0078] Here, (Equation 35) holds.

[0079] Furthermore, in the backpropagation method according to the present disclosure, in addition to the weight matrix obtained by (Equation 34), the phase φ i (l) is given as a parameter and updated. i (l) The update of is given by (Equation 36), similar to (Equation 34).

[0080]

[0081] Here, the partial differential term of the loss function in the second term is given by (Equation 37).

[0082]

[0083] Similar to the error function of (Equation 28), the second term of (Equation 37) is the feedback error from the layer two layers before. Such propagation of the error from the layer two layers before is performed in the pseudo quantum spin layer 103, and in the layer one layer before the output layer 104, the partial differential term of the loss function is given as (Equation 38).

[0084]

[0085] In this way, in the machine learning method according to the present disclosure, error propagation occurs not only from the previous layer but also from the previous two layers. This enables the weight matrix and phase to be updated in a short time. In other words, by using the neural network 100 according to the present disclosure and executing the backpropagation method according to the present disclosure, machine learning that takes quantum mechanical effects into account in a pseudo-manner can be achieved in a short time.

[0086] (Machine Learning Method Specialized for Combinatorial Optimization Problems) Next, as an example of the backpropagation method according to the present disclosure, a form specialized for solving combinatorial optimization problems will be described. However, the combinatorial optimization problem here is limited to a combinatorial optimization problem equivalent to an Ising Hamiltonian minimization problem as described in (Equation 1) or a problem called QUBO (Quadratic Unconstrained Binary Optimization). It is known that the Ising problem and the QUBO problem are applicable to a wide range of practical application problems.

[0087] 5 is a conceptual diagram showing the structure of a neural network 500 according to the present disclosure, which is specialized for solving combinatorial problems. As shown in FIG. 5, the neural network 500 has a structure similar to that of the neural network 100 shown in FIG. 1, but the input layer 501 also functions as a spin layer. The parameters for the backpropagation method are φ i (l) The weight matrix is ​​fixed as shown in (Equation 39).

[0088]

[0089] where f(l) is a function of the number of layers l, and J ij is the interaction matrix of the Ising problem to be solved. The local magnetic field term is expressed by (Equation 40).

[0090]

[0091] Here, this local magnetic field term is a fixed magnetic field that changes as a function of the number of layers 1. For example, it is defined as a mutually increasing and decreasing function, as expressed by (Equation 41) and (Equation 42).

[0092]

[0093] where L is the number of layers in the last layer, and ξ and h x are parameters that adjust the magnitude of the magnetic field and Jij, and appropriate values ​​are set for these.

[0094] Using such equations, we consider solving a combinatorial optimization problem by the backpropagation method according to the present disclosure. First, as an input to the spin layer, xi in = -1, and S yi in , S zi in A minute amount is input to . Then, as with the above-mentioned spin layer, an activation function expressed by (Equation 17) is introduced. The calculations in the intermediate layer (spin layer and pseudo quantum spin layer) are as described above, and (Equation 43) is used as the activation function in the calculation of the final output from the output layer.

[0095]

[0096] Such an operation corresponds to a pseudo-expression of quantum annealing using a neural network. In this neural network, the phase φ i (l) By learning this, it is possible to solve combinatorial optimization problems with high accuracy. The loss function at this time is given by (Equation 44).

[0097]

[0098] Here, U is σ zi L = ±1.

[0099] (Verification of Accuracy of Machine Learning Method) The accuracy of the backpropagation method according to the present disclosure was verified using the neural network 100 according to the present disclosure described above. The details and results of the verification will be described below.

[0100] 6 is a diagram showing the accuracy rates in the MNIST benchmark using the conventional backpropagation method and the backpropagation method according to the present disclosure. The MNIST benchmark is an image of handwritten digits, and includes 60,000 training data sets and 10,000 test data sets. This benchmark is used for learning using the training data, and the accuracy rate of digit recognition is tested using the test data.

[0101] As shown in FIG. 6, it was confirmed that the backpropagation method according to the present disclosure exhibited a higher accuracy rate than the conventional backpropagation method.

[0102] 7 shows the results of predicting time series data in the NARMA10 benchmark using the backpropagation method according to the present disclosure, where (a) shows the target data and predicted data, and (b) shows the error. Note that NARMA10 is time series data given by (Equation 45).

[0103]

[0104] Here, u i is a uniform random number in [0, 1], and the parameters are α=0.3, β=0.05, γ=1.5, and δ=0.1.

[0105] As shown in Figure 7, the predicted data obtained by predicting time series data using the backpropagation method according to the present disclosure showed good agreement with the target data, and it was confirmed that there was almost no error. Predicting time series data is generally a task mainly performed using recurrent neural networks, but the present disclosure achieves high accuracy in such tasks. In other words, it can be said that the backpropagation method according to the present disclosure has high versatility.

[0106] (Example of Hardware Configuration) Figure 8 is a block diagram illustrating the configuration of a computer 800 that implements the machine learning method using a neural network according to the present disclosure. As described above, the computer 800 may be in the form of a general digital computer. As shown in Figure 8, the computer 800 includes an input device 801 that receives input, a processor 802 that executes learning based on the machine learning method using a neural network according to the present disclosure, and an output device 803 that outputs a final output.

[0107] The input device 801 receives inputs (x in FIGS. 3 and 4) sent from a user or a database, etc. in ) and transmits it to the processor 802. The processor 802 performs calculations based on the input received from the input device 801 and the machine learning method according to the present disclosure described above. The final output generated by the processor 802 is transmitted to the output device 803 and output to the outside.

[0108] The computer 800 may further include a memory 804. The memory 804 has a function of temporarily storing input received by the input device 801 and final output output by the processor 802. The computer 800 may also include conversion devices 805a and 805b between the input device 801 and the processor 802, or between the processor 802 and the output device 803. The conversion devices 805a and 805b have a function of converting the input and output into data or signals in a format suitable for the transmission destination.

[0109] As described above, the neural network according to the present disclosure is configured to incorporate pseudo-quantum mechanical effects and use a time evolution equation that approximately represents the quantum annealing method. Furthermore, in a machine learning method using such a neural network, the machine learning method according to the present disclosure is configured to cause error propagation from the two layers prior. Such a neural network and machine learning method according to the present disclosure can achieve high computational efficiency on a general digital computer, and therefore can be applied to machine learning for a variety of tasks.

Claims

1. A neural network comprising: a spin layer including an intermediate layer that includes an integer number of neurons that is a multiple of 3 and generates an output in the form of a spin operator; and a pseudo quantum spin layer that is coupled to the spin layer and includes multiple layers therein; wherein the spin layer generates an output in the form of the spin operator based on (Equation 1) and (Equation 2), and the pseudo quantum spin layer takes into account time evolution based on (Equation 3), and outputs spins based on (Equation 4) and (Equation 5).

2. The neural network of claim 1, wherein the activation function that calculates the final learning output is a soft matrix function.

3. A machine learning method using the neural network described in claim 1, in which the error function is given by (Equation 6), and the weight matrix and phase are updated based on the propagation of errors from the previous layer and the previous two layers.

4. The machine learning method of claim 3, wherein the weight matrix is ​​fixed by (Equation 7), the local magnetic field term is a fixed magnetic field that changes as a function of a sum of 1, the input to the spin layer has an x-component spin of -1, and the y-component spin and z-component spin are infinitesimal amounts, and the activation function in the calculation of the final output is given by (Equation 8).