阅读说明：本技术 积和运算器、神经形态器件以及积和运算器的故障判断方法 (Product-sum calculator, neuromorphic device, and method for determining failure of product-sum calculator ) 是由 柴田龙雄 于 2018-12-12 设计创作，主要内容包括：本发明提供在应用于神经网络的情况下,能够正确地检测可能大幅损坏神经网络的性能的故障的积和运算器。积和运算器(1)具备积运算部(10)、和运算部(11)、故障判断部(12),积运算部(10)具备多个积运算元件(10AA～10AC),多个积运算元件(10AA)～(10AC)各自为电阻变化元件。和运算部(11)具备检测来自多个积运算元件(10AA)～(10AC)的输出的合计值的输出检测器(11A)。故障判断部(12)在输出检测器(11)检测的上述合计值超过规定值的情况下,判断为故障产生。上述规定值是在多个积运算元件(10AA～10AC)全部正常动作的情况下,输出检测器(11)能够检测的上述合计值的最大值以上的值。(The invention provides a product-sum calculator which can accurately detect faults which can greatly damage the performance of a neural network when the product-sum calculator is applied to the neural network. A product-sum calculator (1) is provided with a product calculation unit (10), a sum calculation unit (11), and a failure determination unit (12), wherein the product calculation unit (10) is provided with a plurality of product calculation elements (10AA to 10AC), and each of the plurality of product calculation elements (10AA to 10AC) is a variable resistance element. The sum operation unit (11) is provided with an output detector (11A) that detects the sum of the outputs from the multiple product operation elements (10AA) to (10 AC). A failure determination unit (12) determines that a failure has occurred when the sum detected by the output detector (11) exceeds a predetermined value. The predetermined value is a value equal to or greater than the maximum value of the sum value that can be detected by the output detector (11) when all of the plurality of product computing elements (10AA to 10AC) are operating normally.)

1. A product-sum operator, characterized in that,

comprises an integral calculation unit, a sum calculation unit, and a failure determination unit,

the product operation unit includes a plurality of product operation elements,

each of the plurality of product operation elements is a resistance change element,

the sum operation unit includes an output detector for detecting a total value of outputs from the plurality of product operation elements,

the failure determination unit determines that a failure has occurred when the total value detected by the output detector exceeds a predetermined value,

the predetermined value is a value equal to or greater than a maximum value of the sum value that the output detector can detect when all of the product computing elements are operating normally.

2. The product-sum operator according to claim 1,

the variable resistance element has a write terminal, a common terminal, and a read terminal.

3. The product-sum operator according to claim 1 or 2, wherein,

the resistance change element is a magnetoresistance effect element exhibiting magnetoresistance effect,

the magnetoresistance effect element has:

a magnetization free layer having a magnetic wall;

a magnetization pinned layer whose magnetization direction is pinned; and

a nonmagnetic layer sandwiched between the magnetization free layer and the magnetization pinned layer.

4. The product-sum operator according to any one of claims 1 to 3,

the product operation unit includes at least a first column having a plurality of product operation elements and a second column having a plurality of product operation elements,

the sum operation unit includes at least a first output detector that detects a total value of outputs from the plurality of product operation elements in the first column and a second output detector that detects a total value of outputs from the plurality of product operation elements in the second column,

the failure determination unit determines that the first column has failed, and the function replacement unit causes the plurality of product operation elements of the second column to perform a product operation performed by the plurality of product operation elements of the first column before the failure of the first column.

5. The product-sum operator according to any one of claims 1 to 4,

further provided with:

a failure location specifying unit that specifies a failed product computing element of the plurality of product computing elements when the failure determination unit determines that the plurality of product computing elements are failed;

an input cutoff unit that cuts off an input to the product computing element having the failure when the failure determination unit determines that the plurality of product computing elements have failed,

the failure location specifying unit specifies the failed product computing element by inputting a signal from an output side of the plurality of product computing elements to the plurality of product computing elements.

6. A neuromorphic device provided with the product-sum calculator described in any one of claims 1 to 5.

7. The neuromorphic device of claim 6, wherein,

a hidden layer using an activation function is also provided,

the activation function is set such that an output value from the activation function becomes zero when an input value to the activation function is equal to or greater than a threshold value.

8. A neuromorphic device provided with the product-sum calculator according to any one of claims 1 to 3,

the apparatus further includes an output cutoff unit that cuts off an output from the product elements determined to be defective when the failure determination unit determines that the product elements are defective.

9. The neuromorphic device of any one of claims 6-8, wherein,

the resistance value of each of the plurality of product operation elements in the normal operation differs from the resistance value of each of the plurality of product operation elements in the failure by 3 orders or more.

10. A neuromorphic device provided with the product-sum operator according to claim 1 or 2,

the resistance change element is a magnetoresistance effect element exhibiting magnetoresistance effect,

the magnetoresistance effect element has:

a magnetization free layer having a magnetic wall;

a magnetization pinned layer whose magnetization direction is pinned; and

a nonmagnetic layer sandwiched by the magnetization free layer and the magnetization pinned layer,

the thickness of the nonmagnetic layer is 2.5nm or more.

11. The neuromorphic device of claim 10, wherein,

the magnetization free layer is a perpendicular magnetization film selected from the group consisting of a Co/Pt multilayer film, a Co/Pd multilayer film, and a CoCrPt alloy film.

12. The neuromorphic device of any one of claims 6-11, wherein,

the output detector detects an output current value from the plurality of product operation elements.

13. The neuromorphic device of any one of claims 6-11, wherein,

the output detector detects the electric charges output from the plurality of product operation elements.

14. A failure judgment method for a product-sum operator,

the product-sum calculator comprises a product-sum calculator and a sum calculator,

the product operation unit includes a plurality of product operation elements,

each of the plurality of product operation elements is a resistance change element,

the fault judgment method comprises the following steps:

a detection step of detecting a total value of outputs from the plurality of product operation elements; and

a determination step of determining that a failure has occurred when the total value detected in the detection step exceeds a predetermined value,

the predetermined value is a value equal to or greater than a maximum value of the sum value that can be detected when all of the product computing elements are operating normally.

Technical Field

The invention relates to a product-sum calculator, a neuromorphic device, and a method for determining a failure of a product-sum calculator. The present application claims priority based on patent application No. 2017-254700, which was applied in japan at 12/28/2017, and the contents of which are incorporated herein by reference.

Background

Conventionally, a learning plan for optimization of gray scale image recognition in an RRAM (registered trademark) based neuromorphic system is known (for example, see non-patent document 1). This document describes that a neuromorphic system has been developed based on an assembled resistive switching memory array. In this document, a new training plan is proposed to optimize the performance of the simulation system by using the behavior of the segmented synapses. In this document, the plan is applied to grayscale image recognition.

In addition, studies have been conducted to realize a neural network that mimics the nervous system using an array of resistance change elements. In a neuromorphic device (NMD), a product-sum operation is performed in which weights are applied from a preceding stage to a succeeding stage and accumulated. Therefore, various types of product-sum operators that perform a sum operation by combining a plurality of variable resistance elements whose resistances continuously change, performing a product operation on an input signal with the respective resistance values as weights, and obtaining the sum of currents output from the products have been developed, and NMDs using the product-sum operators have been developed.

Disclosure of Invention

Problems to be solved by the invention

However, non-patent document 1 does not study a method of detecting a failure that may significantly deteriorate the performance of the neural network. If the resistance change element fails and the resistance becomes small, the weight of the failed resistance change element at the time of product sum operation may greatly affect the network. Therefore, it is very important to correctly detect the malfunction of the resistance change element caused by the short circuit in the neural network.

In view of the above-described problems, an object of the present invention is to provide a product-sum calculator, a neuromorphic device, and a method for determining a failure of a product-sum calculator, which are capable of accurately detecting a failure that may significantly deteriorate the performance of a neural network when an element failure occurs, when applied to the neural network.

Means for solving the problems

One aspect of the present invention provides a product-sum calculator including a product calculation unit including a plurality of product calculation elements each of which is a resistance change element, a sum calculation unit including an output detector that detects a total value of outputs from the plurality of product calculation elements, and a failure determination unit that determines that a failure has occurred when the total value detected by the output detector exceeds a predetermined value, the predetermined value being a value equal to or greater than a maximum value of the total value that can be detected by the output detector when all of the plurality of product calculation elements are operating normally.

In the product-sum calculator according to one aspect of the present invention, the variable resistance element may include a write terminal, a common terminal, and a read terminal.

In the product-sum calculator according to one aspect of the present invention, the variable resistance element may be a magnetoresistance effect element that exhibits a magnetoresistance effect, and the magnetoresistance effect element may include: the magnetic recording medium includes a magnetization free layer having a magnetic wall, a magnetization pinned layer having a fixed magnetization direction, and a nonmagnetic layer sandwiched between the magnetization free layer and the magnetization pinned layer.

In the product-sum calculator according to one aspect of the present invention, the product calculation unit may include at least a first column including a plurality of product calculation elements and a second column including a plurality of product calculation elements, the sum calculation unit may include at least a first output detector that detects a sum of outputs from the plurality of product calculation elements in the first column and a second output detector that detects a sum of outputs from the plurality of product calculation elements in the second column, and the sum calculation unit may further include a function replacement unit that, when the failure determination unit determines that the first column has failed, causes the plurality of product calculation elements in the second column to perform a product calculation performed by the plurality of product calculation elements in the first column before the failure of the first column.

The product-sum calculator according to an aspect of the present invention may further include: a failure location specifying unit that specifies a failed product computing element of the plurality of product computing elements when the failure determination unit determines that the plurality of product computing elements are failed; and an input blocking unit that blocks an input to the failed product computing element when the failure determination unit determines that the plurality of product computing elements are failed, wherein the failure location identification unit identifies the failed product computing element by inputting a signal to the plurality of product computing elements from an output side of the plurality of product computing elements.

One embodiment of the present invention provides a neuromorphic device including the product-sum calculator.

The neuromorphic device according to one aspect of the present invention may further include a hidden layer using an activation function set such that an output value from the activation function becomes zero when an input value to the activation function is equal to or greater than a threshold value.

A neuromorphic device according to an aspect of the present invention is a neuromorphic device including the product-sum calculator, and may include an output cutting unit that cuts off an output from the plurality of product calculating elements determined to be defective when the failure determination unit determines that the plurality of product calculating elements are defective.

In the neuromorphic device according to one aspect of the present invention, the resistance value of each of the plurality of product operation elements in the normal operation may be different from the resistance value of each of the plurality of product operation elements in the failure by 3 orders or more.

A neuromorphic device according to an aspect of the present invention may be a neuromorphic device including the product-sum calculator, wherein the variable resistance element is a magnetoresistance effect element that exhibits a magnetoresistance effect, and the magnetoresistance effect element includes: the magnetic recording medium includes a magnetization free layer having a magnetic wall, a magnetization pinned layer having a fixed magnetization direction, and a nonmagnetic layer sandwiched between the magnetization free layer and the magnetization pinned layer, wherein the thickness of the nonmagnetic layer is 2.5nm or more.

In the neuromorphic device according to one embodiment of the present invention, the magnetization free layer may be a perpendicular magnetization film selected from the group consisting of a Co/Pt multilayer film, a Co/Pd multilayer film, and a CoCrPt alloy film.

In the neuromorphic device according to one aspect of the present invention, the output detector may detect an output current value from the plurality of product operation elements.

In the neuromorphic device according to one aspect of the present invention, the output detector may detect the electric charges output from the plurality of product calculating elements.

One aspect of the present invention provides a failure determination method for a product-sum calculator including a product calculation unit and a sum calculation unit, wherein the product calculation unit includes a plurality of product calculation elements each of which is a resistance change element, and the method includes: a detection step of detecting a total value of outputs from the plurality of product operation elements; and a determination step of determining that a failure has occurred when the total value detected in the detection step exceeds a predetermined value, the predetermined value being a value equal to or greater than a maximum value of the total value that can be detected when all of the plurality of product calculators are operating normally.

Effects of the invention

According to the present invention, it is possible to provide a product-sum calculator, a neuromorphic device, and a method for determining a failure of a product-sum calculator, which can accurately detect a failure that may significantly deteriorate the performance of a neural network when applied to the neural network.

Drawings

Fig. 1 is a diagram showing an example of a partial configuration of a product-sum calculator according to a first embodiment;

fig. 2 is a diagram showing an example of the overall configuration of the product-sum calculator according to the first embodiment;

fig. 3 is a diagram for explaining the total value and the predetermined value detected by the output detector;

fig. 4 is a perspective view showing an example of a magnetoresistive element constituting each of a plurality of product operation elements of the product-sum calculator according to the first embodiment;

fig. 5 is a diagram showing an application example of the product-sum calculator according to the first embodiment;

fig. 6 is a diagram for explaining an activation function used in a hidden layer;

fig. 7 is a flowchart showing an example of processing executed by the product-sum calculator according to the first embodiment;

fig. 8 is a diagram showing an example of a partial configuration of the product-sum calculator according to the second embodiment.

Detailed Description

Hereinafter, embodiments of a product-sum calculator, a neuromorphic device, and a method for determining a failure in a product-sum calculator according to the present invention will be described with reference to the drawings.

< first embodiment > (the resistance change element is a magnetoresistance effect element)

Fig. 1 is a diagram showing an example of a partial configuration of a product-sum calculator 1 according to the first embodiment. Fig. 2 is a diagram showing an example of the overall configuration of the product-sum calculator 1 according to the first embodiment.

In the example shown in fig. 1 and 2, the product-sum calculator 1 according to the first embodiment includes: the product calculation unit 10, the sum calculation unit 11, the failure determination unit 12, the function replacement unit 13, the failure site identification unit 14, the input disconnection unit 15, and the output disconnection unit 16. The product operation unit 10 includes columns 10A and 10B.

In the example shown in fig. 1, the product calculation unit 10 includes two columns 10A and 10B, but may include only one column, or may include 3 or more columns 10A, 10B, and ….

In the example shown in fig. 1, the column 10A includes an product operation element 10AA, an product operation element 10AB, and a product operation element 10 AC. The column 10B includes an integration element 10BA, an integration element 10BB, and an integration element 10 BC.

In the example shown in fig. 1, the column 10A includes 3 product computing elements 10AA, 10AB, and 10AC, and the column 10B includes 3 product computing elements 10BA, 10BB, and 10BC, but the column 10A may include a plurality of product computing elements other than 3, and the column 10B may include a plurality of product computing elements other than 3.

In the example shown in fig. 1, each of the plurality of product computing elements 10AA to 10AC and 10BA to 10BC is a variable resistance element having a read terminal, a write terminal, and a common terminal.

The read terminals of the product elements 10AA and 10BA are connected to a line L11, and the write terminals of the product elements 10AA and 10BA are connected to a line L12.

The read terminals of the product elements 10AB and 10BB are connected to a line L21, and the write terminals of the product elements 10AB and 10BB are connected to a line L22.

The read terminals of the product elements 10AC, 10BC are connected to the line L31, and the write terminals of the product elements 10AC, 10BC are connected to the line L32.

The common terminals of the product operation elements 10AA, 10AB, and 10AC are connected to a line M1. The common terminal of the product elements 10BA, 10BB, 10BC is connected to the line M2.

The sum operation unit 11 includes: an output detector 11A for detecting the total value of the outputs from the product operation elements 10AA, 10AB, and 10 AC; and an output detector 11B for detecting a total value of outputs from the product computing elements 10BA, 10BB, and 10 BC. The output detector 11A is disposed on the line M1. The output detector 11B is disposed on the line M2.

In the example shown in fig. 1, the output detector 11A detects output current values from the product elements 10AA, 10AB, and 10AC, and the output detector 11B detects output current values from the product elements 10BA, 10BB, and 10 BC. In another example, the output detector 11A detects the outputs from the product elements 10AA, 10AB, and 10AC as charges, and the output detector 11B detects the outputs from the product elements 10BA, 10BB, and 10BC as charges.

The resistance variable element used as the product operation element in the present invention is an element in which the resistance reversibly changes with respect to a stimulus (current, voltage, magnetic field, or the like) from the outside. Examples of the variable resistance element include: resistance change memory (RRAM) elements, phase change memory (PCRAM) elements, Anisotropic Magnetoresistance (AMR) elements, Tunnel Magnetoresistance (TMR) elements, Giant Magnetoresistance (GMR) elements, and the like.

When the total value detected by the output detector 11A exceeds a predetermined value, the failure determination unit 12 determines that a failure has occurred in which the output current increases in at least one of the plurality of product operation elements 10AA to 10AC included in the column 10A. When the total value detected by the output detector 11B exceeds a predetermined value, the failure determination unit 12 determines that a failure has occurred in at least one of the plurality of product operation elements 10BA to 10BC included in the column 10B, the failure having a large output current.

Fig. 3 is a diagram for explaining the total value and the predetermined value detected by the output detector 11A. In fig. 3, the vertical axis represents the total value, predetermined value, and the like of the output currents of the multiple product computing elements 10AA to 10AC detected by the output detector 11A. The horizontal axis represents the states (the magnitudes of the resistance values) of the product operation elements 10AA to 10 AC.

When the plurality of product computing elements 10AA to 10AC are in normal operation and the resistance values of the product computing elements 10AA to 10AC are the highest, the total value of the output currents of the plurality of product computing elements 10AA to 10AC detected by the output detector 11A becomes the minimum value Min.

During normal operation of the plurality of product computing elements 10AA to 10AC, the total value of the output currents of the plurality of product computing elements 10AA to 10AC detected by the output detector 11A increases as the resistance values of the product computing elements 10AA to 10AC decrease.

When the plurality of product computing elements 10AA to 10AC are in normal operation and the resistance values of the product computing elements 10AA to 10AC are the lowest, the total value of the output currents of the plurality of product computing elements 10AA to 10AC detected by the output detector 11A becomes the maximum value Max.

The predetermined value is set to a value not less than the maximum value Max. That is, the predetermined value is a value equal to or larger than the maximum value Max of the total values that can be detected by the output detector 11A when all of the plurality of product computing elements 10AA to 10AC are operating normally.

In the example shown in fig. 3, in the state of the point P1, the total value of the output currents of the product operation elements 10AA to 10AC detected by the output detector 11A does not exceed a predetermined value. Therefore, the failure determination unit 12 determines that no failure having an increased output current occurs in any of the plurality of product operation elements 10AA to 10AC included in the column 10A.

On the other hand, in the state of the point P2, the total value of the output currents of the product elements 10AA to 10AC detected by the output detector 11A exceeds a predetermined value. Therefore, the failure determination unit 12 determines that a failure in which the output current increases occurs in at least one of the plurality of product operation elements 10AA to 10AC included in the column 10A.

Specifically, in the examples shown in fig. 1 and 3, the product elements 10AA to 10AC (and the product elements 10BA to 10BC) are magnetoresistive effect elements each showing a magnetoresistive effect. The resistance values of the plurality of product computing elements 10AA to 10AC in the normal operation and the resistance values of the plurality of product computing elements 10AA to 10AC in the failure may be different by 3 orders or more.

Fig. 4 is a perspective view showing an example of the magnetoresistive element a constituting each of the plurality of product operation elements 10AA to 10AC and 10BA to 10BC of the product-sum calculator 1 according to the first embodiment.

In the example shown in fig. 4, the magnetoresistance effect element a includes: a magnetization free layer a1 having a magnetic wall DW, a magnetization pinned layer a2 in which the magnetization direction is pinned, and a nonmagnetic layer A3. The nonmagnetic layer A3 is sandwiched by the magnetization free layer a1 and the magnetization pinned layer a 2. The magnetization free layer a1 has a first region a11 on one side of the magnetic wall DW and a second region a12 on the other side of the magnetic wall DW. The first area a11 is provided with a write terminal AA. The second region a12 is provided with a common terminal AB. The magnetization pinned layer a2 has a read terminal AC provided therein.

The amount of movement (distance of movement) of the magnetic wall DW can be controlled to be variable by adjusting the magnitude and time of the write current flowing between the write terminal AA and the common terminal AB. The magnitude and time of the write current can also be set according to, for example, the number of pulses or the pulse width, to set the moving amount (moving distance) of the magnetic wall DW. When the areas of the magnetization parallel (or antiparallel) portions of the magnetization pinned layer a2 and the magnetization free layer a1 are continuously changed by the drive (movement) of the magnetic wall DW, the ratio of the area ratio of the magnetization parallel portion to the area ratio of the magnetization antiparallel portion is continuously changed, and a nearly linear resistance change is obtained in the magnetoresistance effect element.

Data can be read by passing a current between the read terminal AC and the common terminal AB and detecting a resistance according to the ratio of the area ratio of the portion parallel to the magnetization direction to the area ratio of the portion antiparallel to the magnetization direction (see, for example, patent document 1).

[ magnetization pinned layer A2]

The magnetization pinned layer a2 is a layer in which magnetization is oriented in a first direction (for example, the left direction in fig. 4) and pinned. Here, the fixed magnetization means that the magnetization direction does not change (the magnetization is fixed) before and after writing using a write current.

In the example shown in fig. 4, the magnetization pinned layer a2 is an in-plane magnetization film whose magnetization has in-plane magnetic anisotropy (in-plane easy magnetization axis). The magnetization pinned layer a2 is not limited to the in-plane magnetization film, and may be a perpendicular magnetization film having perpendicular magnetic anisotropy (perpendicular easy magnetization axis).

When the magnetization pinned layer a2 is an in-plane magnetization film, it has a high MR ratio (magnetoresistance change rate), and is less susceptible to the influence of Spin Transfer Torque (STT) at the time of reading, and thus the read voltage can be increased. On the other hand, when the element is miniaturized, it is preferable to use a perpendicular magnetization film having a large magnetic anisotropy and a small diamagnetic field.

A known material can be used for the magnetization pinned layer a 2. For example, a metal selected from the group consisting of Cr, Mn, Co, Fe, and Ni, and an alloy containing 1 or more of these metals and exhibiting ferromagnetism can be used. Further, an alloy containing at least 1 or more of these metals, B, C, and N can also be used. Specifically, Co-Fe and Co-Fe-B are mentioned.

Co can also be used for the magnetization pinned layer A2₂A heusler alloy of FeSi, etc. The heusler alloy comprises a metal having X₂YZ, X is a transition metal element or a noble metal element of Co, Fe, Ni or Cu group in the periodic table, Y can be a transition metal of Mn, V, Cr or Ti group, or can adopt the element type of X, and Z is a typical element of III-V group. Examples thereof include: co₂FeSi、Co₂MnSi、Co₂Mn_1-aFe_aAl_bSi_1-bAnd the like.

The magnetization pinned layer a2 may have a composite structure including a ferromagnetic layer and a nonmagnetic layer, or a composite structure including an antiferromagnetic layer, a ferromagnetic layer, and a nonmagnetic layer. In the latter, in the resultant structure, the magnetization direction of the magnetization pinned layer a2 is strongly held by the antiferromagnetic layer. Therefore, the magnetization of the magnetization pinned layer a2 is hardly affected by external factors.

When the magnetization of the magnetization pinned layer a2 is oriented in the XY plane (the magnetization pinned layer a2 is an in-plane magnetization film), NiFe is preferably used, for example. On the other hand, when the magnetization of the magnetization pinned layer a2 is oriented in the Z direction (the magnetization pinned layer a2 is a perpendicular magnetization film), for example, a Co/Ni laminated film, a Co/Pt laminated film, or the like is preferably used. For example, when the magnetization pinned layer A2 is set to [ Co (0.24nm)/Pt (0.16nm) ]]₆/Ru(0.9nm)/[Pt(0.16nm)/Co(0.16nm)]₄and/Ta (0.2nm)/FeB (1.0nm), a perpendicular magnetization film is formed.

[ nonmagnetic layer A3]

The nonmagnetic layer A3 is disposed on the lower surface of the magnetization pinned layer a 2. The magnetoresistance effect element a reads out a change in the magnetization state of the magnetization free layer a1 with respect to the magnetization fixed layer a2 as a change in resistance value via the nonmagnetic layer A3. That is, the magnetization pinned layer a2, the nonmagnetic layer A3, and the magnetization free layer a1 function as the magnetoresistance effect element a, and have a structure similar to a Tunnel Magnetoresistance (TMR) element when the nonmagnetic layer A3 is made of an insulator, and a structure similar to a Giant Magnetoresistance (GMR) element when the nonmagnetic layer 2 is made of a metal.

As the material of the nonmagnetic layer a3, a known material that can be used for the nonmagnetic layer of the magnetoresistive element a can be used. When the nonmagnetic layer a3 is made of an insulator (in the case of a tunnel barrier layer), the following materials can be used: al (Al)₂O₃、SiO₂、MgO、MgAl₂O₄、ZnAl₂O₄、MgGa₂O₄、ZnGa₂O₄、MgIn₂O₄、ZnIn₂O₄And multilayer films or mixed composition films of these materials. In addition to these, materials in which a part of Al, Si, and Mg is replaced with Zn, Be, and the like can Be used. Of these, MgO and MgAl₂O₄The material that can realize coherent tunneling is used, and thus the magnetoresistance ratio (MR ratio) can be increased. On the other hand, when the nonmagnetic layer 2 is made of metal, Cu, Al, Ag, or the like can be used as a material thereof.

When the nonmagnetic layer a3 is made of an insulator (in the case of a tunnel barrier layer), the thickness thereof is, for example, 2.5nm or more.

[ magnetization free layer A1]

The magnetization free layer a1 corresponds to a magnetic wall drive layer of a magnetic wall drive type (mobile type) MRAM.

The magnetization free layer a1 is made of a ferromagnetic material, and the direction of magnetization therein can be reversed. The magnetization free layer a1 has: a first region a11 in which the magnetization is oriented in a second direction opposite to the magnetization pinned layer a2, a second region a12 in which the magnetization is oriented in the same direction as the first direction, and a magnetic wall DW constituting an interface between these regions. The first region a11 sandwiching the magnetic wall DW has the opposite direction of magnetization to the second region a 12. The magnetic wall DW is moved by the change in the composition ratio of the first region a11 and the second region a12 of the magnetization free layer a 1.

As the material of the magnetization free layer a1, a known material can be used, and in particular, a soft magnetic material can be used. For example, a metal selected from the group consisting of Cr, Mn, Co, Fe, and Ni, an alloy containing 1 or more of these metals, an alloy containing B, C and at least 1 or more of these metals and N, and the like can be used. Specifically, examples of the material of the magnetization free layer A1 include Co-Fe and Co-Fe-B, Ni-Fe.

The material of the magnetization free layer a1 can also be a material with a small saturation magnetization. For example, when a material having a small saturation magnetization, such As (MnGa) As and (InFe) As, or a Co/Tb multilayer film and GdFeCo, is used, the magnetic wall DW of the magnetization free layer a1 can be driven at a small current density. In addition, when these materials are used, the driving speed of the magnetic wall DW becomes slow.

The magnetic wall DW of a material with weak magnetic anisotropy, such as NiFe, is driven at a high speed, and the magnetic wall DW operates at a speed of 100m/sec or more. That is, the magnetic wall DW moves a distance of 1 μm in a pulse of 10 nsec. Therefore, when the magnetization free layer a1 is moved in an analog manner within the element, it is necessary to take measures such as applying a minute pulse using an expensive semiconductor circuit or extending the magnetization free layer sufficiently at the expense of integration. In contrast, in the case of a material in which the driving speed of the magnetic wall DW is slow, an analog memory can be formed even in the case where a sufficiently long pulse current is applied or the case where the length of the magnetization free layer a1 is short.

When the magnetization free layer a1 is a perpendicular magnetization film, it is preferably a perpendicular magnetization film selected from the group consisting of a Co/Pt multilayer film, a Co/Pd multilayer film, and a CoCrPt alloy film. In addition, Mn can also be used₃Perpendicular magnetization films of X (Ga, Ge) or multilayer films of Co/Ni, etc. These materials can drive the magnetic wall DW even if the current density for magnetic wall driving is small.

The length of the magnetization free layer a1 extending in the longitudinal direction is preferably 60nm or more. If the thickness is less than 60nm, a single magnetic domain is likely to be formed, and it is difficult to form the magnetic wall DW in the magnetization free layer A1.

The thickness of the magnetization free layer a1 is not particularly limited as long as it functions as a magnetic wall drive layer, and can be set to 2nm to 60nm, for example. When the thickness of the magnetization free layer a1 is 60nm or more, the possibility of forming a magnetic wall along the stacking direction becomes high. However, whether or not the magnetic wall is formed along the stacking direction is generated by the balance with the shape anisotropy of the magnetization free layer a 1. If the thickness of the magnetization free layer a1 is less than 60nm, it is considered that it is difficult to form the magnetic wall DW.

The magnetization free layer a1 may also have a magnetic wall pinning portion that blocks the movement of the magnetic wall DW on the side of the layer. For example, when the irregularities, grooves, ridges, constrictions, notches, and the like are provided at positions where the movement of the magnetic wall DW of the magnetization free layer a1 is to be blocked, the movement (pinning) of the magnetic wall can be blocked. When the magnetic wall pinning portion is provided, if a current equal to or larger than a threshold value is not passed, the magnetic wall can be configured not to move more than the threshold value, and the output signal is not an analog signal, which facilitates multi-valued conversion.

For example, by forming the magnetic wall pinning portion every predetermined distance, the magnetic wall DW can be held more stably, stable multi-value recording can be performed, and multi-valued output signals can be read more stably.

In the example shown in fig. 4, in order to form the magnetic wall DW, the magnetization free layer a1 has a first magnetization supply layer a4 having the same magnetization in the first direction as the magnetization of the first region a11 and a second magnetization supply layer a5 having the same magnetization in the second direction as the magnetization of the second region a12 at both ends which do not overlap with the magnetization fixed layer a2 in a plan view.

As the material of the first magnetization supplying layer a4 and the second magnetization supplying layer a5, the same ferromagnetic material as that usable for the magnetization pinned layer a2 can be used.

In the example shown in fig. 4, the first magnetization supply layer a4 and the second magnetization supply layer a5 are used as layers for fixing magnetization at both ends of the magnetization free layer a1 in order to form the magnetic wall DW, but Spin Orbit Torque (SOT) wiring that is in contact with the magnetization free layer a1 and extends in a direction intersecting the longitudinal direction of the magnetization free layer a1 may be used for either one or both of them. The spin orbit torque wiring is made of a material which generates a pure spin current by a spin hall effect when a current flows.

With the above-described configuration, even if the magnetization supply layer as a layer for fixing magnetization is not provided, the magnetic wall can be introduced into the magnetization free layer a1 by passing a current to both ends of the spin orbit torque wiring, and the magnetic wall can be moved by passing a current to the magnetization free layer a1 via the spin orbit torque wiring.

In the example shown in fig. 4, the first magnetization supply layer a4 and the second magnetization supply layer a5 are used as layers for fixing magnetization at both ends of the magnetization free layer a1 in order to form the magnetic wall DW, but magnetic field applying wirings that are electrically insulated from the magnetization free layer a1 and extend in a direction intersecting the magnetization free layer a1 may be used for either one or both of them. By applying a current to the magnetic field application wire, a magnetic field is generated according to ampere's law. The direction of the generated magnetic field can be reversed depending on the direction of the current flowing through the magnetic field application wiring. Therefore, by arranging the in-plane magnetization at the end of the magnetization free layer a1 so as to be suppliable, it is possible to supply magnetization in one of the in-plane magnetization directions opposite to each other to the end of the magnetization free layer a1 in accordance with the direction of current flowing through the magnetic field application wiring. In addition, by disposing the perpendicular magnetization at the end of the magnetization free layer a1 so as to be suppliable, it is possible to supply magnetization in one of the perpendicular magnetization directions opposite to each other to the end of the magnetization free layer a1 in accordance with the direction of current flowing through the magnetic field application wiring.

In the example shown in fig. 4, the first magnetization supply layer a4 and the second magnetization supply layer a5 are used as layers for fixing magnetization at both ends of the magnetization free layer a1 in order to form the magnetic wall DW, but a voltage application terminal connected to the magnetization free layer a1 via an insulating layer may be used for either or both of them. When a voltage is applied between the magnetization fixed layer a2 and the voltage application terminal, a part of the magnetization free layer a1 is affected by the voltage. For example, when a voltage is applied in pulses from a voltage application terminal, a part of the magnetization is oriented in a direction orthogonal to the magnetization direction of the magnetization free layer a1 when the voltage is applied, and the magnetization of the magnetization free layer a1 is oriented in a second direction that is the first direction or the opposite direction thereto when the voltage application is stopped. The probability that the magnetization oriented in the orthogonal direction is tilted in the first direction or in the second direction opposite thereto is equal, and by adjusting the timing, the number of times, and the period of applying the pulse voltage, it is possible to orient a part of the magnetization from the first direction to the second direction.

A magnetic coupling layer may also be provided between the magnetization free layer a1 and the nonmagnetic layer A3. The magnetic coupling layer is a layer that transfers the magnetization state of the magnetization free layer a 1. The main function of the magnetization free layer a1 is a layer for driving the magnetic wall, and is not limited to being able to select a material suitable for the magnetoresistance effect generated via the magnetization fixed layer a1 and the nonmagnetic layer a 2. It is generally known that, in order to generate a coherent tunneling effect using the nonmagnetic layer a2, a ferromagnetic material of BCC structure of the magnetization pinned layer a1 and the magnetic coupling layer is good. In particular, it is known that a material having a composition of Co-Fe-B as the material of the magnetization pinned layer A1 and the magnetic coupling layer provides a large output when manufactured by sputtering.

In the example shown in fig. 1 and 2, when the failure determination unit 12 determines that the column 10A has failed, the function replacement unit 13 causes the plurality of product operation elements 10BA to 10BC in the column 10B to perform product operation of the plurality of product operation elements 10AA to 10AC in the column 10A before the failure of the column 10A. Specifically, for example, the function replacement unit 13 sets the resistance values of the plurality of product operation elements 10BA to 10BC in the column 10B to the resistance values of the plurality of product operation elements 10AA to 10AC in the column 10A before failure.

When the failure determination unit 12 determines that the column 10B has failed, the function replacement unit 13 causes the plurality of product operation elements 10AA to 10AC of the column 10A to perform product operation of the plurality of product operation elements 10BA to 10BC of the column 10B before the failure of the column 10B. Specifically, for example, the function replacement unit 13 sets the resistance values of the plurality of product operation elements 10AA to 10AC in the column 10A to the resistance values of the plurality of product operation elements 10BA to 10BC in the column 10B before failure.

For example, when the failure determination unit 12 determines that any one of the plurality of product computing elements 10AA to 10AC in the column 10A has failed, the failure location determination unit 14 determines which one of the plurality of product computing elements 10AA to 10AC has failed. Specifically, the failure-point specifying unit 14 specifies which product computing element fails by inputting signals to the plurality of product computing elements 10AA to 10AC from the output side (lower side in fig. 1) of the plurality of product computing elements 10AA to 10 AC.

When the failure determination unit 12 determines that any one of the plurality of product operation elements 10AA to 10AC in the column 10A has failed, for example, when the failure location determination unit 14 determines that the product operation element 10AB has failed, the input disconnection unit 15 disconnects the input to the failed product operation element 10 AB. As a result, it is possible to avoid an excessive current from being output to the line M1 via the product operation element 10 AB.

In one example of the input disconnecting unit 15 for disconnecting the input to the product operation element 10AB, the input disconnecting unit 15 is configured by a switch disposed between the readout terminal AC (see fig. 4) of the product operation element 10AB and the line L21, and in another example of the input disconnecting unit 15 for disconnecting the input to the product operation element 10AB, the input disconnecting unit 15 is configured by a switch for disconnecting the input to the product operation elements 10AB and 10BB on the line L21.

Instead of the input cutoff unit 15 cutting off the input to the failed product computing element 10AB, the output cutoff unit 16 may cut off the outputs from the plurality of product computing elements 10AA to 10AC including the failed product computing element 10AB on, for example, a line M1.

Similarly, when the failure determination unit 12 determines that any one of the plurality of product elements 10BA to 10BC in the column 10B has failed, the failure site specification unit 14 specifies which one of the plurality of product elements 10BA to 10BC has failed. Specifically, the failure site specifying unit 14 specifies which product element fails by inputting a signal to the plurality of product elements 10BA to 10BC from the output side (lower side in fig. 1) of the plurality of product elements 10BA to 10 BC.

When the failure determination unit 12 determines that any one of the plurality of product operation elements 10BA to 10BC in the column 10B has failed, for example, when the failure site specification unit 14 specifies that the product operation element 10BC has failed, the input disconnection unit 15 disconnects the input to the failed product operation element 10 BC. As a result, it is possible to avoid an excessive current from being output to the line M2 via the product element 10 BC.

Instead of the input cutoff unit 15 cutting off the input to the failed product computing element 10BC, the output cutoff unit 16 may cut off the outputs from the plurality of product computing elements 10BA to 10BC including the failed product computing element 10BC on, for example, a line M2.

Fig. 5 is a diagram showing an application example of the product-sum calculator 1 according to the first embodiment.

In the example shown in fig. 5, the product-sum calculator 1 of the first embodiment is applied to a neuromorphic device 100. The neuromorphic device 100 includes: an input layer 101, a hidden layer 102, an output layer 103, a product-sum operator 1 of the first embodiment, and a product-sum operator 2. The product-sum operator 2 has a plurality of product-operation elements, as in the product-sum operator 1 of the first embodiment shown in fig. 1.

The input layer 101 includes, for example, 4 nodes 101A, 101B, 101C, and 101D. The hidden layer 102 includes, for example, 3 nodes 102A, 102B, and 102C. The output layer 103 includes, for example, 3 nodes 103A, 103B, and 103C.

The product-sum operator 1 is disposed between the input layer 101 and the hidden layer 102, and connects 4 nodes 101A, 101B, 101C, and 101D of the input layer 101 to 3 nodes 102A, 102B, and 102C of the hidden layer 102. The product-sum calculator 1 changes the weights by changing the resistance values of the product calculating elements 10AA to 10AC and 10BA to 10BC shown in fig. 1.

A product-sum calculator 2 is disposed between the hidden layer 102 and the output layer 103. The product-sum operator 2 connects the 3 nodes 102A, 102B, 102C of the hidden layer 102 with the 3 nodes 103A, 103B, 103C of the output layer 103. The product-sum calculator 2 changes the weight by changing the resistance values of the plurality of product-operation elements.

The hidden layer 102 uses an activation function (e.g., Sigmoid function).

Fig. 6 is a diagram for explaining an activation function used in the hidden layer 102. In fig. 6, the horizontal axis represents the input value to the activation function, and the vertical axis represents the output value from the activation function.

In the example shown in fig. 6, the activation function is set such that the output value from the activation function becomes zero when the input value to the activation function is equal to or greater than the threshold value.

That is, in the example shown in fig. 5 and 6, for example, when a value equal to or greater than a threshold value (see fig. 6) is input from the product-sum calculator 1 to the hidden layer 102 due to a failure of the product-sum calculator 1, the hidden layer 102 outputs a value of zero as an output value. Therefore, for example, the influence of a failure of the product-sum calculator 1 or the like can be suppressed from affecting the possible rows of the product-sum calculator 2 and the output layer 103.

Fig. 7 is a flowchart showing an example of processing performed by the product-sum calculator 1 according to the first embodiment.

In step S10, the product operation unit 10 and the sum operation unit 11 perform product-sum operation.

In step S11, the output detector 11A detects the output current values from the product elements 10AA, 10AB, and 10AC, and the output detector 11B detects the output current values from the product elements 10BA, 10BB, and 10 BC.

In step S12, failure determination unit 12 determines whether or not the total value detected by output detector 11A exceeds a predetermined value, and determines whether or not the total value detected by output detector 11B exceeds a predetermined value. When the total value detected by the output detector 11A exceeds the predetermined value or when the total value detected by the output detector 11B exceeds the predetermined value, the process proceeds to step S13. When the total value detected by the output detector 11A is equal to or less than a predetermined value and the total value detected by the output detector 11B is equal to or less than a predetermined value, the processing shown in fig. 7 is ended.

In step S13, the failure determination unit 12 determines that a failure has occurred in which the output current increases in at least one of the plurality of product operation elements 10AA to 10AC included in the column 10A, or determines that a failure has occurred in which the output current increases in at least one of the plurality of product operation elements 10BA to 10BC included in the column 10B.

The present inventors have found, through extensive studies, that the function of the neuromorphic device 100 is degraded when the characteristics of the product computing elements 10AA to 10AC and 10BA to 10BC constituting the product-sum computing unit 1 are changed by some factors (specifically, when the product computing elements 10AA to 10AC and 10BA to 10BC fail).

In particular, it has been found that when the product operation elements 10AA to 10AC and 10BA to 10BC fail in the short circuit mode (that is, when a failure in which the output current increases occurs in the product operation elements 10AA to 10AC and 10BA to 10BC), the product operation function of the product operation unit 1 and the performance as a neural network are significantly impaired.

Specifically, the present inventors have found, through extensive studies, that when a failure in which the output current increases occurs in the product operation elements 10AA to 10AC and 10BA to 10BC, the failure in which the output current decreases is greater than when the failure occurs in the product operation elements 10AA to 10AC and 10BA to 10BC, and the product-sum operation function of the product-sum operator 1 and the performance as a neural network are damaged to a greater extent. This is because a large amount of current flows intensively in a defective product operation element (resistance variable element), and thus the weight of another product operation element (current from another product operation element) becomes invisible on the circuit.

Therefore, in the product-sum calculator 1 according to the first embodiment, as described above, the failure determination unit 12 determines that a failure has occurred in any one of the plurality of product calculating elements 10AA to 10AC as the variable resistance elements when the total value detected by the output detector 11A exceeds a predetermined value, and determines that a failure has occurred in any one of the plurality of product calculating elements 10BA to 10BC as the variable resistance elements when the total value detected by the output detector 11B exceeds a predetermined value. The predetermined value is a value equal to or greater than the maximum value of the total value that can be detected by the output detector 11A when all of the plurality of product computing elements 10AA to 10AC are operating normally, and is a value equal to or greater than the maximum value of the total value that can be detected by the output detector 11B when all of the plurality of product computing elements 10BA to 10BC are operating normally.

Therefore, according to the product-sum operator 1 of the first embodiment, in the case of being applied to a neural network, it is possible to accurately detect a failure that may significantly deteriorate the performance of the neural network (that is, a failure in which the output current is large).

In the product-sum calculator 1 of the first embodiment, as described above, each of the plurality of product calculating elements 10AA to 10AC and 10BA to 10BC is a magnetoresistive element a having a write terminal AA, a common terminal AB, and a read terminal AC and exhibiting a magnetoresistive effect. Further, the magnetoresistance effect element a includes: a magnetization free layer a1 having a magnetic wall DW, a magnetization pinned layer a2 in which the magnetization direction is pinned, and a nonmagnetic layer A3 sandwiched between the magnetization free layer a1 and the magnetization pinned layer a 2.

That is, in the product-sum calculator 1 according to the first embodiment, elements having a large difference between the resistance value in the normal operation and the resistance value in the failure (more specifically, in the failure in which the output current is large) are used as the plurality of product calculating elements 10AA to 10AC, 10BA to 10 BC.

Therefore, according to the product-sum calculator 1 of the first embodiment, it is possible to accurately detect a failure that may significantly deteriorate the performance of the neural network, as compared to a case where an element having a small difference between the resistance value in the normal operation and the resistance value in the failure (in detail, in the failure in which the output current is large) is used.

In the product-sum calculator 1 according to the first embodiment, when the failure determination unit 12 determines that the column 10A has failed, the product calculation performed by the plurality of product calculating elements 10AA to 10AC of the column 10A before the failure of the column 10A is performed by the plurality of product calculating elements 10BA to 10BC of the column 10B, as described above. When the failure determination unit 12 determines that the column 10B has failed, the plurality of product operation elements 10BA to 10BC of the column 10B are used to perform product operation on the plurality of product operation elements 10AA to 10AC of the column 10A before the failure of the column 10B.

Therefore, according to the product-sum calculator 1 of the first embodiment, even when the column 10A fails or when the column 10B fails, the performance of the neural network can be maintained.

In the product-sum calculator 1 according to the first embodiment, when the failure determination unit 12 determines that any one of the plurality of product calculating elements 10AA to 10AC has failed as described above, the failed product calculating element of the plurality of product calculating elements 10AA to 10AC is identified, and the input to the failed product calculating element is cut off.

When the failure determination unit 12 determines that any one of the plurality of product elements 10BA to 10BC has failed, the failed product element of the plurality of product elements 10BA to 10BC is specified, and the input to the failed product element is cut off.

Therefore, according to the product-sum calculator 1 of the first embodiment, it is possible to suppress the possibility of significantly deteriorating the performance of the neural network, compared to the case where the input to the failed product calculating element is not cut off.

Alternatively, in the product-sum calculator 1 according to the first embodiment, when the failure determination unit 12 determines that any one of the plurality of product operation elements 10AA to 10AC has failed, the output from the plurality of product operation elements 10AA to 10AC determined to have failed is cut off as described above.

When the failure determination unit 12 determines that any one of the plurality of product elements 10BA to 10BC has failed, the output from the plurality of product elements 10BA to 10BC determined to have failed is cut off.

Therefore, according to the product-sum calculator 1 of the first embodiment, it is possible to suppress the possibility of significantly impairing the performance of the neural network, as compared with the case where the outputs from the plurality of product calculating elements 10AA to 10AC, 10BA to 10BC determined to be faulty are not cut off.

In the neuromorphic device 100 including the product-sum calculator 1 according to the first embodiment, as described above, the activation function used in the hidden layer 102 is set such that the output value from the activation function becomes zero when the input value to the activation function is equal to or greater than the threshold value.

Therefore, according to the neuromorphic device 100 including the product-sum calculator 1 of the first embodiment, for example, a failure of the product-sum calculator 1 or the like can suppress the influence on the product-sum calculator 2 and the output layer 103.

< second embodiment > (the resistance variable element is a common variable resistor)

A second embodiment of the product-sum calculator of the present invention will be described below.

The product-sum calculator 1 of the second embodiment is configured in the same manner as the product-sum calculator 1 of the first embodiment, except for the points described later. Therefore, according to the product-sum calculator 1 of the second embodiment, the same effects as those of the product-sum calculator 1 of the first embodiment described above can be achieved except for the points described below.

Fig. 8 is a diagram showing an example of a partial configuration of the product-sum calculator 1 according to the second embodiment.

In the product-sum calculator 1 of the first embodiment, the product calculating elements 10AA to 10AC and 10BA to 10BC are formed by magnetoresistive elements, but in the product-sum calculator 1 of the second embodiment, the product calculating elements 10AA to 10AC and 10BA to 10BC are formed by ordinary variable resistance elements (variable resistors).

Specifically, in the example shown in fig. 1, each of the plurality of product computing elements 10AA to 10AC and 10BA to 10BC includes a read terminal, a write terminal, and a common terminal, but in the example shown in fig. 8, each of the plurality of product computing elements 10AA to 10AC and 10BA to 10BC includes a first terminal and a second terminal.

In the example shown in fig. 8, the first terminals of the product operation elements 10AA and 10BA are connected to a line L11, the first terminals of the product operation elements 10AB and 10BB are connected to a line L21, and the first terminals of the product operation elements 10AC and 10BC are connected to a line L31.

Second terminals of the product operation elements 10AA, 10AB, 10AC are connected to a line M1. Second terminals of the product elements 10BA, 10BB, 10BC are connected to the line M2.

The sum operation unit 11 includes: an output detector 11A for detecting a total value of outputs from the second terminals of the product computing elements 10AA, 10AB, and 10 AC; and an output detector 11B for detecting a total value of outputs from the second terminals of the product computing elements 10BA, 10BB, and 10 BC.

While the embodiments of the present invention have been described in detail with reference to the drawings, the specific configuration is not limited to the embodiments, and various modifications and substitutions can be made without departing from the spirit of the present invention. The configurations described in the above embodiments may be combined.

For example, the processing may be performed by recording a program for realizing the functions of each device (for example, product-sum calculator 1) of the above-described embodiments in a computer-readable recording medium (storage medium), and causing a computer system to read and execute the program recorded in the recording medium.

The "computer System" may include hardware such as an Operating System (OS) and peripheral devices.

The "computer-readable recording medium" refers to a storage device such as a flexible disk, an optical disk, a writable nonvolatile memory such as a rom (read Only memory) or a flash memory, a portable medium such as a dvd (digital Versatile disc), or a hard disk incorporated in a computer system. The recording medium may be a recording medium on which data is temporarily recorded, for example.

The "computer-readable recording medium" also includes a medium that holds a program for a certain period of time, such as a volatile memory (for example, dram (dynamic Random Access memory)) inside a computer system serving as a server or a client when the program is transmitted via a network such as the internet or a communication line such as a telephone line.

The program may be transmitted from a computer system storing the program in a storage device or the like to another computer system via a transmission medium or by a transmission wave in the transmission medium. Here, the "transmission medium" for transmitting the program refers to a medium having a function of transmitting information, such as a network (communication network) such as the internet or a communication line (communication line) such as a telephone line.

The program may be a program for realizing a part of the above-described functions. The program may be a so-called differential file (differential program) that can realize the above-described functions in combination with a program already recorded in a computer system.

In a computer, a processor such as a cpu (central Processing unit) reads out and executes a program stored in a memory.

Description of the symbols

1A 1 product-sum arithmetic unit, a2 product-sum arithmetic unit, a 10 product arithmetic unit, a 10A column, a 10AA product arithmetic element, a 10AB product arithmetic element, a 10AC product arithmetic element, a 10B column, a 10BA product arithmetic element, a 10BB product arithmetic element, a 10BC product arithmetic element, an 11 sum arithmetic unit, an 11A output detector, an 11B output detector, a12 failure determination unit, a 13 function replacement unit, a 14 failure location determination unit, a 15 input cutoff unit, a 16 output cutoff unit, a 100 neuromorphic device, a 101 input layer, a 101A, 101B, 101C, 101D node, a 102 hidden layer, 102A, 102B, 102C node, 103 output layer, 103A, 103B, 103C node, a magnetoresistance effect element, an AA write terminal, an AB common terminal, an AC read terminal, an a magnetization free layer, a first region, a second region, a magnetization fixed layer, a nonmagnetic layer, magnetic wall, 11 line, 12 line, 21 line, 22 line, 31, M line.