WO2018212201A1

WO2018212201A1 - Information processing device and information processing method

Info

Publication number: WO2018212201A1
Application number: PCT/JP2018/018809
Authority: WO
Inventors: 野村　光; 義茂鈴木; 田村　英一; 真嗣三輪; 穣後藤; ペパーフェルディナンド
Original assignee: 国立大学法人大阪大学; 国立研究開発法人情報通信研究機構
Priority date: 2017-05-15
Filing date: 2018-05-15
Publication date: 2018-11-22
Also published as: JP7108987B2; JPWO2018212201A1

Abstract

This information processing device comprises: a magnetic reservoir that includes a plurality of nodes, each of the nodes being made of a magnetic body, and magnetically coupled to at least one other node; a node selection unit that selects one or more nodes to which information is to be written from among the plurality of nodes included in the magnetic reservoir; an information writing unit that writes information to the one or more selected nodes; and an information reading unit that, after the information is written, reads information attained by linear coupling of each node included in the magnetic reservoir.

Description

Information processing apparatus and information processing method

The present invention relates to an information processing apparatus and an information processing method.

Currently, many of the computers that are widely used are Neumann computers. The Neumann computer includes a memory for storing programs and data, and an information processing device. The information processing device provided in the von Neumann computer executes appropriate arithmetic processing according to programs and data stored in the memory.

On the other hand, there are non-Neumann computers. As one of non-Neumann computers, a reservoir computer imitating a neural network has been proposed (see, for example, Patent Documents 1 and 2).

US Patent Application Publication No. 2014/0214738 U.S. Pat. No. 9,477,136

However, in the reservoir computer disclosed in Patent Document 1, the operation node constituting the reservoir is composed of negative resistance elements, and information is written by current control on each negative resistance element. Further, even when information is held in a reservoir provided with a negative resistance element as an operation node, current control is required. Therefore, the reservoir computer disclosed in Patent Document 1 has a problem that power consumption cannot be kept low.

Further, in the reservoir computer disclosed in Patent Document 2, the computation node constituting the reservoir is configured by a passive silicon photonics chip, and writing of information requires irradiation with laser light or coherent light. Therefore, the reservoir computer disclosed in Patent Document 2 also has a problem that power consumption cannot be kept low.

An object of the present invention is to provide an information processing apparatus and an information processing method capable of realizing reservoir computing with low power consumption.

An information processing apparatus according to an aspect of the present invention includes a magnetic material reservoir that includes a plurality of nodes, each of which is made of a magnetic material, each of which is magnetically coupled to at least one other node, and the magnetic material A node selection unit for selecting one or more nodes to write information from a plurality of nodes included in the body reservoir, an information writing unit for writing information to the selected one or more nodes, and after writing the information And an information reading unit for reading information obtained by linear combination of each node included in the magnetic substance reservoir.

An information processing method according to one aspect of the present invention includes a plurality of nodes each formed of a magnetic material, and each node is magnetically coupled to at least one other node. Select one or more nodes to which information is to be written, write information to the selected one or more nodes, and after writing the information, read information obtained by linear combination of each node contained in the magnetic reservoir .

According to the above aspect, reservoir computing can be realized with low power consumption.

It is a conceptual diagram which shows the structure of the information processing apparatus which concerns on this Embodiment. It is a schematic diagram which shows an example of a micro magnetic body. It is a flowchart which shows the procedure of the process which the information processing apparatus which concerns on this Embodiment performs. It is a schematic diagram which shows the structure of the micro magnetic substance reservoir used by performance evaluation. 6 is a schematic diagram showing a configuration of a minute magnetic substance reservoir 20 according to Embodiment 2. FIG. It is a graph explaining the change method of magnetic anisotropy. It is explanatory drawing explaining the data used for a binary calculation. It is a graph which shows the error rate of each calculation at the time of learning using the Z component of the magnetization state between time 0-1. It is a graph which shows the error rate of each calculation at the time of learning using the Z component of the magnetization state between time 0-1. It is a graph which shows the error rate of each calculation at the time of learning using the Z component of the magnetization state between time 0-1. 6 is a graph showing an error rate of each calculation when learning is performed using an X component of a magnetization state between

times

1 and 2; 6 is a graph showing an error rate of each calculation when learning is performed using an X component of a magnetization state between

times

1 and 2; It is a graph which shows the error rate with respect to the delay amount of input. FIG. 10 is a schematic diagram illustrating a configuration of a minute magnetic substance reservoir according to a third embodiment. It is explanatory drawing explaining the change method of magnetic anisotropy. It is a figure which shows the comparison result with teacher data. It is a figure which shows the comparison result with teacher data. It is a graph which shows the error rate in the exclusive logic calculation at the time of learning using the X component of a magnetization state. FIG. 10 is a schematic diagram showing a configuration of a minute magnetic substance reservoir according to a fourth embodiment. FIG. 10 is a schematic plan view of a minute magnetic substance reservoir according to a fifth embodiment. FIG. 10 is a schematic cross-sectional view of a minute magnetic substance reservoir according to a fifth embodiment.

Hereinafter, the present invention will be specifically described with reference to the drawings showing embodiments thereof.
(Embodiment 1)
FIG. 1 is a conceptual diagram showing the configuration of the information processing apparatus according to the present embodiment. The information processing apparatus according to the present embodiment is a so-called reservoir computer, and includes an input information cell 10, a minute magnetic substance reservoir 20, a weighted arithmetic element 30, and an output information cell 40.

The micro magnetic material reservoir 20 includes a plurality of

operation nodes

21, 21,. Each operation node 21 is composed of a minute magnetic material to be described later. Since the magnetostatic interaction acts between the minute magnetic bodies, the

operation nodes

21 and 21 are magnetically coupled. In the example of FIG. 1, among the 4 rows × 5 columns of

operation nodes

21, 21,..., 21 arranged in a matrix, the operation node 21 arranged in the upper left corner is adjacent to the two on the right side and the lower side. It shows that the

operation nodes

21 and 21 are magnetically coupled. The same applies to the other operation nodes 21. In the example of FIG. 1, each operation node 21 is magnetically coupled to 2 to 4 other operation nodes 21.

The example of FIG. 1 shows a state in which the

nearest computation nodes

21 and 21 are magnetically coupled to each other, but the two

computation nodes

21 and 21 that are magnetically coupled are not necessarily arranged closest to each other. There is no need to be done.

In the example of FIG. 1, the configuration in which the

operation nodes

21, 21,..., 21 are arranged in a matrix of 4 rows × 5 columns is shown, but the arrangement of the operation nodes 21 is limited to the example of FIG. is not. For example, as shown in FIG. 1, an operation node 21 may be arranged at each vertex of a square lattice, and an operation node 21 is arranged at each vertex of various planar lattices such as a rectangular lattice, a triangular lattice, a hexagonal lattice, and a rhombus lattice. May be. Further, the operation nodes 21 do not have to be arranged on a plane, and may be arranged at each vertex in an arbitrary space lattice. Further, the operation nodes 21 do not have to be periodically arranged, and may be randomly arranged in the same plane or space.

The micro magnetic material constituting the computation node 21 is formed of an alloy such as a Ni—Fe alloy, a Ni—Fe—Co alloy, a Co—Fe alloy, or the like. The micromagnetic material is configured so as to have only one easy axis using, for example, shape magnetic anisotropy. FIG. 2 is a schematic view showing an example of a minute magnetic material. The micro magnetic material used for the computation node 21 has, for example, an elliptic cylinder shape, and has a size of about 100 nm in the major axis direction of the ellipse, 50 nm in the minor axis direction, and about 20 nm in the thickness direction. In such a micro magnetic material, only one easy axis is formed in a direction that coincides with the major axis direction of the ellipse. In particular, if the minute magnetic body is elongated along the easy axis, the direction of magnetization exhibits bistability on the easy axis. That is, the magnetization direction in the micromagnetic material is the first direction along the easy magnetization axis (the direction indicated by the white arrow in the figure), or the second direction reversed 180 degrees from the first direction (the black arrow in the figure). It can be configured such that any one of the directions (indicated by a symbol) is taken and the other direction (the direction deviated from the easy axis of magnetization) is not taken.

In FIG. 2, as an example of the micro magnetic material, the configuration of the micro magnetic material having only one easy magnetization axis and showing the bistability of the magnetization direction on the easy magnetization axis has been described. As long as the stable direction of magnetization is determined according to the state, it may be a micromagnetic material that does not exhibit bistability, and may be a micromagnetic material that does not have an easy magnetization axis or a micromagnetic material that has multiple easy magnetization axes. There may be. In addition, in FIG. 2, the minute magnetic body having an elliptic cylinder shape has been described, but the shape of the minute magnetic body is not limited to the elliptic cylinder shape. For example, the shape of the cross section orthogonal to the thickness direction may be a circle, a rectangle, a rectangle with rounded corners, or a shape obtained by slightly overlapping two circles. Micromagnetic materials having these shapes are easy to manufacture and have the advantage of easy integration.

The input information cell 10 writes information to the

operation nodes

21, 21,..., 21 included in the minute magnetic substance reservoir 20. The input information cell 10 writes information to one or more selected

operation nodes

21, 21,..., 21 among the plurality of

operation nodes

21, 21,. . Information to be written by the input information cell is described by binary values of 0 or 1, for example. When writing “0” to the selected computation node 21, the input information cell 10 controls the magnetization direction of the micromagnetic material constituting the computation node 21 to the first direction, for example, and writes “1”. The magnetization direction of the minute magnetic material is controlled to a second direction that is 180 degrees reversed from the first direction.

As a method for controlling the magnetization direction in a micro magnetic material, for example, a method used in a magnetoresistive effect type solid magnetic memory (MRAM: Magnetic Random Access Memory) can be used. As a method for controlling the magnetization direction used in the MRAM, a method using a classic current magnetic field or a method using spin injection magnetization reversal is known. In the former, it is possible to generate a magnetic field by flowing a current through a wiring arranged in the very vicinity of the target micromagnetic material, and to control the magnetization direction by the generated magnetic field. In the latter, magnetization reversal can be performed by supplying polarized spin current (spin injection), and the magnetization direction of the target micromagnetic material can be controlled. However, in the former, there is a demerit that the required current increases as the size of the magnetic material decreases, and a write error may occur in another micro magnetic material close to the target micro magnetic material due to a leakage magnetic field from the wiring. There are disadvantages. For this reason, in the present embodiment, it is preferable to use the latter method of spin injection magnetization reversal.

Further, instead of the classical current magnetic field method or the spin injection magnetization reversal method, a method of inducing a magnetic response by the electric field effect may be used. For example, in International Publication No. 2009/133650, an electric field is applied to a structure in which a magnetic layer (ultra-thin ferromagnetic layer) and an insulating layer (potential barrier) are stacked, so that the magnetic layer A technique for modulating the magnetic anisotropy and controlling the magnetization direction of the magnetic layer is disclosed. Such a method may be used to control the magnetization direction of the minute magnetic body constituting the selected computation node 21.

After the information is written to the selected

computation nodes

21, 21,..., 21 by the above-described magnetization control method, the magnetization direction of the minute magnetic body constituting each computation node 21 is minute. Under the influence of magnetostatic interaction from the magnetic material, the state autonomously transits to a state representing the calculation result.

In the output information cell 40, after information is written by the input information cell 10, the timing at which the magnetization state of the minute magnetic substance reservoir 20 transitions to a state representing the calculation result (for example, timing at which 10 nsec has elapsed after writing information). Read information. Specifically, the output information cell 40 linearly combines the information held in the

calculation nodes

21, 21,..., 21 of the minute magnetic substance reservoir 20 through the weighting calculation element 30. Read the information obtained by.
For reading out information from each of the

operation nodes

21, 21,..., 21, it is possible to use a technique used in the MRAM as in the magnetization direction control technique described above.

The weighting calculation element 30 has a function of linearly combining information held in the

calculation nodes

21, 21,..., 21 of the minute magnetic substance reservoir 20, and a function of determining a linear weight used in the linear combination by learning. . When the output instruction is given from the outside, the weighting calculation element 30 reads the linear weight stored in the memory not shown in the figure, and uses the read linear weight to each of the

calculation nodes

21, 21,. The stored information is linearly combined, and the obtained information is output to the output information cell 40. The weighting calculation element 30 may hold the weight by a memory element made of semiconductor, or may hold the weight by a magnetic memory using the position of the domain wall in the magnetic body.

On the other hand, when a learning instruction for linear weight is given from the outside, the weighting calculation element 30 reads teacher information indicating an ideal output with respect to the input from a memory not shown in the figure, and reproduces the teacher information. The linear weight is determined by learning. The weighting calculation element 30 stores the determined linear weight in the memory described above. The weighting calculation element 30 uses, for example, a least square method (linear regression) so that the residual sum of squares between the value indicated by the information obtained from the minute magnetic substance reservoir 20 and the value indicated by the teacher information is minimized. Linear weights may be determined. Further, when the number of operation nodes 21 included in the minute magnetic substance reservoir 20 is too large, linear weights may be regularized by Ridge regression, Lasso regression, Elastic net method, or the like in order to avoid problems due to overlearning. . Furthermore, it is possible to perform the learning process by real-time processing instead of badge processing by using a recursive least square method or a so-called FORCE learning method.

In the present embodiment, information is written from the input information cell 10 to the minute magnetic material reservoir 20, but a plurality of

input nodes

11, 11,..., 11 are provided in the input information cell 10, and the input nodes 11, .., 11 may be used to write information to the

operation nodes

21, 21,. In addition, a plurality of

output nodes

41, 41,..., 41 may be provided in the output information cell 40, and information may be read from the minute magnetic substance reservoir 20 using each of the

output nodes

41, 41,.

Hereinafter, the operation of the information processing apparatus according to the present embodiment will be described.
FIG. 3 is a flowchart showing a procedure of processing executed by the information processing apparatus according to this embodiment. The input information cell 10 of the information processing apparatus selects one or a plurality of

operation nodes

21, 21,..., 21 to which information is to be written when the operation target information is input and an operation instruction is given (step S101). . Here, the input information cell 10 may select at least 10% to 90% of the

operation nodes

21, 21,..., 21 out of all the

operation nodes

21, 21,.

For example, when only the

computation nodes

21, 21,..., 21 arranged in a specific row (or column) are selected from the computation nodes 21 shown in FIG. 1 and information is written, these

computation nodes

21, 21,. , 21 may stabilize the magnetization direction. In this case, the written information does not propagate to the

other computation nodes

21, 21,..., 21 and cannot function as a reservoir. Therefore, it is preferable that the

operation nodes

21, 21,..., 21 selected in step S101 are random. However, it is not always necessary to randomly select the

operation nodes

21, 21,..., 21 as long as information can be periodically input so as to be in a high energy state, and the operation node 21 is given periodicity. , 21,..., 21 may be selected.

Next, the input information cell 10 writes information to one or a plurality of

operation nodes

21, 21,..., 21 selected in step S101 (step S102). The input information cell 10 writes information by controlling the magnetization direction of the minute magnetic body constituting each operation node 21. For controlling the magnetization direction, a method using a classic current magnetic field, a method using spin injection magnetization reversal, a method of inducing a magnetic response by the electric field effect, or the like can be used.

Next, the magnetization state of the minute magnetic substance reservoir 20 is changed to a state representing the calculation result (step S103). The magnetization direction of the micro magnetic material constituting the calculation node 21 is affected by the magnetostatic interaction from the micro magnetic material arranged around the calculation node 21 and autonomously transits to a state representing the calculation result. . As a process executed by the information processing apparatus, a process of waiting for a time (for example, 10 nsec) required for the magnetization direction of the minute magnetic substance reservoir 20 to transition to a state representing a calculation result is performed.

Next, the output information cell 40 controls and outputs the information held in the minute magnetic substance reservoir 20 (step S104). Specifically, the output information cell 40 linearly combines the information held in each of the

operation nodes

21, 21,... 21 via the weighted operation element 30 and outputs information obtained by the linear combination. Read as information. If there is further input information, the information processing apparatus reads the output information, returns the process to step S102, and continues the calculation.

In addition, although the sequential procedure was shown in the flowchart of FIG. 3, a plurality of sets of steps S101 to S104 may be prepared and these sets may be executed asynchronously.

Hereinafter, performance evaluation of the information processing apparatus according to the present embodiment will be disclosed.
FIG. 4 is a schematic diagram showing the configuration of the minute magnetic substance reservoir 20 used in the performance evaluation. The micro magnetic material reservoir 20 shown in FIG. 4 has 16 × 16 square lattices (unit cells 200), the first micro magnetic material 201 at the apex of each square lattice, and the second at the center of each square lattice. The micro magnetic body 202 is arranged. That is, in the example of FIG. 4, two micro

magnetic bodies

201 and 202 are arranged in one unit cell 200, and a micro magnetic reservoir 20 having 16 × 16 × 2 operation nodes 21 is constructed as a whole. ing. In this example, the magnetostatic interaction that works between the first micromagnetic body 201, the magnetostatic interaction that works between the first micromagnetic body 201 and the second micromagnetic body 202, and the second micromagnetic body. The magnetostatic interaction acting between 202 is used for computation.

In the example shown in FIG. 4, when writing 16-bit information, out of 16 × 16 × 2 operation nodes 21, for example, 16 operation nodes 21 (3.125% of all operation nodes 21). Information can be written by selecting at random.

In this embodiment, in the small magnetic material reservoir 20 shown in FIG. 4, 30% of the

computation nodes

21, 21,..., 21 are selected at random, and the NARMA 10 (NonlinearonAuto-Regressive Moving Average 10) task is selected. The performance was evaluated by executing In addition, the magnetization direction of the micro magnetic material in each operation node 21 was calculated by solving the Landau-Lifshitz equation using the fourth-order Runge-Kutta method.

According to the calculation results by the inventors, the normalized error value (NRMSE: normalized root mean square error) in the NARMA10 task shows a value of 0.8 or less, and the micromagnetic material reservoir according to the present embodiment It has been found that 20 can be used for reservoir computing.

(Embodiment 2)
In the second embodiment, a verification result of the minute magnetic substance reservoir 20 in which the arrangement of the operation nodes 21 is different from that in the first embodiment will be described.

FIG. 5 is a schematic diagram showing the configuration of the minute magnetic substance reservoir 20 according to the second embodiment. The micro magnetic material reservoir 20 shown in FIG. 5 has an Nth row (N is an integer of 1 to 10), such as one computation node 21 in the first row, two

computation nodes

21, 21,. , 21 are arranged in N. Each computation node 21 is composed of a columnar minute magnetic body having a film thickness of 0.5 nm and a radius of 20 nm, and the distance between adjacent minute magnetic bodies is 10 nm. In the present embodiment, the operation nodes 21 are grouped in units of rows, and the operation nodes 21 arranged in the second row, the fifth row, and the eighth row are grouped in the

groups

1, 3, 6, The operation node 21 arranged in the ninth row is group 2, and the operation node 21 arranged in the first, fourth, seventh, tenth and tenth rows is group 3.

The information processing apparatus according to the present embodiment writes information to the operation node 21 arranged in the first row through the input information cell 10 and shifts the information in the minute magnetic material reservoir 20 to each operation node 21. The magnetic anisotropy of the micro magnetic material constituting the structure was changed. The magnetic anisotropy in the micromagnetic material includes shape magnetic anisotropy, interface magnetic anisotropy, crystal magnetic anisotropy, and the like. The information processing apparatus, for example, applies an electric field in the stacking direction to a structure in which a magnetic layer (ultra-thin ferromagnetic layer) and an insulating layer (potential barrier) are stacked, thereby magnetic anisotropy of the micromagnetic material. Can be changed.

FIG. 6 is a graph illustrating a method for changing the magnetic anisotropy. In the graph shown in FIG. 6, the horizontal axis represents time, and the vertical axis represents magnetic anisotropy, and shows the time change of magnetic anisotropy in the minute magnetic bodies belonging to each group 1 to 3. . In order to avoid complication of the graph, +2 is added to the magnetic anisotropy of the micromagnetic material belonging to group 2, and +4 is added to the magnetic anisotropy of the micromagnetic material belonging to group 3. .

In the time step from time 0 to 1, the magnetic anisotropy of the minute magnetic materials belonging to all groups is set to “1”. In this case, the information written in the first line is not shifted to the other operation node 21. However, since a magnetic interaction acts between the operation nodes 21, the magnetization state is affected by the influence, and the stabilization direction is determined.

By effectively bringing the magnetic anisotropy not caused by the magnetic coupling between the

operation nodes

21 and 21 close to “0”, the magnetization direction of the operation node 21 is affected by the magnetic interaction with the surrounding nodes. The direction of magnetization is changed greatly from the state before the magnetic anisotropy is changed.

In the subsequent time step of time 1-2, with the magnetic anisotropy of the micromagnetic material belonging to group 3 set to “1”, the magnetic anisotropy of the micromagnetic material belonging to group 1 and group 2 is set to “0”. Therefore, the information written in the first row can be shifted to the operation node 21 in the second row belonging to the group 1 and the operation node 21 in the third row belonging to the group 2. The same applies to the subsequent time steps of time 2 to 7, and the information held by the operation node 21 in the third row belonging to group 2 in the time step of time 3 to 4 A shift to the operation node 21 can be made. In addition, information held in the operation node 21 in the fourth row belonging to the group 3 can be shifted to the operation node 21 in the fifth row belonging to the group 1 in the time step of time 5 to 6. In this way, by changing the magnetic anisotropy at each time step, it is possible to propagate the information input to the operation node 21 in the first row to each operation node 21. It is assumed that the magnetic anisotropy change rule is stored in advance in a memory not shown in the figure, and the information processing apparatus periodically performs magnetic anisotropy with respect to time according to the change rule stored in the memory. It shall change gender.

Also, by changing the magnetic anisotropy from the time 0 to the time 6 state, the minute magnetic substance reservoir 20 is changed to a state where the next information can be input.

In the second embodiment, the performance of the minute magnetic substance reservoir 20 is evaluated by performing a binary operation using a time-dependent input signal. As binary operations, three types of logical product (AND), logical sum (OR), and exclusive logical sum (XOR) were used. FIG. 7 is an explanatory diagram for explaining data used for binary calculation. A binary state of 0 or 1 is used for input. The binary state input is set to 0 or 1 by setting the polarity of the Z component of the magnetization of the input node (the operation node 21 in the first row shown in FIG. 5) to + or −. The values A and B shown in FIG. 7 are used for the binary calculation that the minute magnetic substance reservoir 20 learns. That is, assuming that the value input at time n _step is A, the data input in the past by n _d steps from A is B.

In the second embodiment, the calculation error rate with respect to the input delay amount in the case where the minute magnetic substance reservoir 20 is learned to perform binary calculation with respect to different inputs is obtained.

8A to 8C are graphs showing the error rate of each calculation when learning is performed using the Z component of the magnetization state between time 0 and time 1. FIG. In the graphs shown in FIGS. 8A to 8C, the horizontal axis represents the delay amount of the input A with respect to the input B, and the vertical axis represents the magnetization of the computation node 21 normalized by the saturation magnetization used for learning of the minute magnetic substance reservoir 20. Significant number of orientations. 8A to 8C show error rates of the logical product operation, the logical sum operation, and the exclusive logical sum operation, respectively. In the logical product operation and the logical sum operation, when the delay amount is less than 4, the error rate is substantially zero. On the other hand, in the exclusive OR operation, the operation can be learned with a relatively large significant number, but in other cases, the error rate is approximately 0.5 or more, indicating that the operation cannot be sufficiently performed.

FIG. 9A to FIG. 9C are graphs showing the error rate of each calculation when learning is performed using the X component of the magnetization state between time 1 and time 2. In the graphs shown in FIGS. 9A to 9C, the horizontal axis represents the delay amount of the input A with respect to the input B, and the vertical axis represents the magnetization of the computation node 21 normalized by the saturation magnetization used for learning of the minute magnetic substance reservoir 20. Significant number of orientations. 9A to 9C show error rates of the logical product operation, the logical sum operation, and the exclusive logical sum operation, respectively. In the logical product operation and the logical sum operation, when the delay amount is less than 4, the error rate is substantially zero. In addition, in the exclusive logic operation, the error rate is kept low when the delay amount is less than 4, and the operation can be learned with the least significant number when the X component of the magnetization between the

time

1 and 2 is used. I found out.

FIG. 10 is a graph showing an error rate with respect to an input delay amount. The graph of FIG. 10 shows the input delay time in the case where the minute magnetic substance reservoir 20 is learned to perform the logical product operation, the logical sum operation, and the exclusive logical sum operation with respect to the input at different times. Indicates the error rate. The horizontal axis of the graph represents the amount of delay n _d, and the vertical axis represents the error rate. As is apparent from the graph of FIG. 10, the error rate is substantially zero for data with a delay amount n _{d up} to 3 for each of the logical product operation, logical sum operation, and exclusive OR operation. Was confirmed.

From the above results, it is clear that when a micro magnetic element is used as a reservoir computer, a logical product operation, a logical sum operation, and an exclusive logical sum operation can be performed using information for the past three pieces. It was. Further, by inverting the polarity of the learned output matrix, it is possible to realize a negative logical product (NAND), negative logical sum (NOR), or negative exclusive logical sum (XNOR) gate.

In the present embodiment, an arrangement in which the operation nodes 21 are increased by one in the direction of the column formed by the group is used. However, an arrangement having a certain number of operation nodes 21 in the column direction may be used. . For example, an arrangement having one or two operation nodes 21 in the direction of a column formed by a group may be used.

In the present embodiment, the configuration in which the groups of the operation nodes 21 are arranged for 10 rows is shown, but the number of rows is not limited to 10 rows. For example, it is possible to increase the amount of delay that can be calculated by increasing the number of rows.

(Embodiment 3)
In the third embodiment, a verification result of the minute magnetic substance reservoir 20 in which two

operation nodes

21 and 21 are arranged in each row will be described.

FIG. 11 is a schematic diagram showing a configuration of the minute magnetic substance reservoir 20 according to the third embodiment. The minute magnetic substance reservoir 20 shown in FIG. 11 has a configuration in which two

operation nodes

21 and 21 are arranged in each of the first to Nth rows (N = 10 in the example shown in FIG. 11). Each computation node 21 is composed of a columnar minute magnetic body having a film thickness of 0.5 nm and a radius of 20 nm, and the distance between adjacent minute magnetic bodies is 10 nm. In the present embodiment, the operation nodes 21 are grouped in units of rows, and the operation nodes 21 arranged in the second row, the fifth row, and the eighth row are grouped in the

groups

The information processing apparatus according to the present embodiment writes information to one of the operation nodes 21 arranged in the first row through the input information cell 10 and shifts the information in the minute magnetic material reservoir 20 with each operation. The magnetic anisotropy of the minute magnetic body constituting the node 21 was changed. The magnetic anisotropy in the micromagnetic material includes shape magnetic anisotropy, interface magnetic anisotropy, crystal magnetic anisotropy, and the like. The information processing apparatus, for example, applies an electric field in the stacking direction to a structure in which a magnetic layer (ultra-thin ferromagnetic layer) and an insulating layer (potential barrier) are stacked, thereby magnetic anisotropy of the micromagnetic material. Can be changed.

FIG. 12 is an explanatory diagram for explaining a method of changing the magnetic anisotropy. In FIG. 12, a black circle indicates a state in which the magnetic anisotropy of the minute magnetic material is adjusted to “1” (a state where K _u ≠ 0), and a white circle indicates a state in which the magnetic anisotropy of the minute magnetic material is adjusted to “0” ( K _u = 0 state). The magnetic anisotropy is adjusted, for example, by applying a voltage to the minute magnetic material. The minute magnetic body indicated by a white circle shows a state in which the magnetic anisotropy is adjusted to “0” in the present embodiment, but the magnetic anisotropy does not have to be completely zero, and has a value close to zero ( Ku≈0).

In the time step represented by Stage 0, the magnetic anisotropy of the minute magnetic materials belonging to all groups is set to “1”. In this case, the information written in the operation node 21 is not shifted to another operation node 21. However, since a magnetic interaction acts between the operation nodes 21, the magnetization state is affected by the influence, and the stabilization direction is determined.

In the subsequent time step represented by Stage 1, the magnetic anisotropy of the micromagnetic materials belonging to Group 1 and Group 2 is set to “0” in a state where the magnetic anisotropy of the micromagnetic materials belonging to Group 3 is set to “1”. Therefore, the information written in the first row can be shifted to the operation node 21 in the second row belonging to the group 1 and the operation node 21 in the third row belonging to the group 2. The same applies to the time steps represented by subsequent Stages 2 to 6, and the information held by the operation node 21 in the third row belonging to Group 2 in the time step in Stage 4 is the operation in the fourth row belonging to Group 3. Shift to node 21. In addition, in the time step of Stage 5, the information held by the computation node 21 in the fourth row belonging to the group 3 can be shifted to the computation node 21 in the fifth row belonging to the group 1. In this way, by changing the magnetic anisotropy at each time step, it is possible to propagate the information input to the operation node 21 in the first row to each operation node 21. It is assumed that the magnetic anisotropy change rule is stored in advance in a memory not shown in the figure, and the information processing apparatus periodically performs magnetic anisotropy with respect to time according to the change rule stored in the memory. It shall change gender.

Further, by changing the magnetic anisotropy from the Stage 0 to the Stage 6 state, the minute magnetic substance reservoir 20 is changed to a state where the next information can be input.

In the third embodiment, the performance of the minute magnetic material reservoir 20 is evaluated by performing a binary operation using a time-dependent input signal. Specifically, the exclusive OR XOR (u _k , u _k-nd ) for the current input u _k and the n _d previous input u _k-nd is trained as a teacher function, which is different from the training time. The performance of the minute magnetic substance reservoir 20 was evaluated by confirming the reproducibility of the teacher data when using the input.

In the training of the minute magnetic reservoir 20, the teacher data _{y k (= XOR (u k} , u k-nd)) minimize obtain closest output o _k to the weighting matrix W two to state M _k of the reservoir at each time step Calculated using multiplication. In this embodiment, training data for 742 type steps is used.

∙ Using the input different from that at the time of training, the output obtained by the weighted matrix W that has been trained was binarized and compared with the teacher data. In the present embodiment, test data for 247 time steps is used.

13A and 13B are diagrams showing comparison results with teacher data. FIG. 13A shows the output data obtained from the response of the micromagnetic material reservoir 20 when new data different from the time of training is input using the weighting matrix W obtained by the pre-training. The vertical axis represents the state of “0” or “1”, and the horizontal axis represents time (step). In this example, an exclusive OR XOR (u _k , u _k-2 ) for the current input u _k and the previous input u _k-2 (ie, n _d = 2) is trained as a teacher function. Shows the results. It can be seen that the output result shown in FIG. 13A completely matches the teacher data shown in FIG. 13B.

FIG. 14 is a graph showing an error rate in the exclusive logical operation when learning is performed using the X component of the magnetization state. In the graph shown in FIG. 14, the horizontal axis represents the delay amount n _d , and the vertical axis represents the effective number of the magnetization direction of the operation node 21 normalized by the saturation magnetization used for learning of the minute magnetic substance reservoir 20. Yes. From the graph shown in FIG. 14, it can be seen that the circuit operates as an XOR function with three or more significant digits and a delay of three.

As described above, in the present embodiment, when a micro magnetic element is used as a reservoir computer, it has become clear that an exclusive OR operation can be performed using information for the past three. Further, a negative exclusive OR (XNOR) gate can be realized by reversing the polarity of the learned output matrix.

(Embodiment 4)
In the fourth embodiment, the configuration of the minute magnetic substance reservoir 20 in which one operation node 21 is arranged in each row will be described.

FIG. 15 is a schematic diagram showing the configuration of the minute magnetic substance reservoir 20 according to the fourth embodiment. The micro magnetic material reservoir 20 shown in FIG. 15 has a configuration in which one operation node 21 is arranged in each of the first to Nth rows (N = 10 in the example shown in FIG. 15). Each computation node 21 is composed of a columnar minute magnetic body having a film thickness of 0.5 nm and a radius of 20 nm, and the distance between adjacent minute magnetic bodies is 10 nm. In the present embodiment, the operation nodes 21 are grouped in units of rows, and the operation nodes 21 arranged in the second row, the fifth row, and the eighth row are grouped in the

groups

According to the study by the inventors, even the minute magnetic substance reservoir 20 in which one operation node 21 is arranged in each row as shown in FIG. 15 can function as an operation gate for exclusive OR or the like. I understood that I could do it.

(Embodiment 5)
In the first to fourth embodiments, the configuration of the minute magnetic substance reservoir 20 in which the calculation nodes 21 are arranged in a lattice shape has been described. However, the calculation nodes 21 do not necessarily need to be periodically arranged, and may be within the same plane or space. May be arranged randomly.
In the fifth embodiment, the configuration of the minute magnetic substance reservoir 20 in which the operation nodes 21 are randomly arranged will be described.

FIG. 16 is a schematic plan view of the minute magnetic substance reservoir 20 according to the fifth embodiment, and FIG. 17 is a schematic sectional view thereof. 16 includes a first layer 20A in which a plurality of

operation nodes

21, 21,..., 21 are arranged, and a second layer 20B in which a node 22 for inputting information is arranged. Have. The second layer 20B is disposed adjacent to the upper side of the first layer 20A, for example. The micro magnetic material constituting each of the

nodes

21 and 22 is formed of an alloy such as a Ni—Fe alloy, a Ni—Fe—Co alloy, or a Co—Fe alloy. Although the minute magnetic body which comprises each

node

21 and 22 has comprised the elliptical column shape, for example, it is not limited to this. For example, the minute magnetic body constituting each of the

nodes

21 and 22 may have a cross-sectional shape that is orthogonal to the thickness direction, a circular shape, a rectangular shape, a rectangular shape with rounded corners, or a slightly overlapping shape of two circles. It may be a shape such as a spheroid.

The first layer 20A has, for example, 6 × 5 unit cells. .., 21 are randomly arranged in each unit cell. The number and arrangement of the operation nodes 21 may be different between units or may be the same.

1 or a plurality of

nodes

22, 22, ..., 22 are formed in the second layer 20B. The node 22 may be formed so as to overlap with one of the

operation nodes

21, 21,... 21 in a plan view, or may be formed so as not to overlap.

The information written in the node 22 of the second layer 20B is propagated to the operation node 21 of the first layer 20A under the influence of the magnetic interaction as in the first to fifth embodiments. In addition, the magnetization direction of each micro magnetic material constituting the

operation nodes

21, 21,... Transition to the state.

As described above, in the fifth embodiment, since the layer including the operation node 21 and the layer including the node 22 to which information is to be written can be manufactured individually, the ease of manufacturing can be improved.

As described above, in the information processing apparatus according to the present embodiment, since the physical properties of the magnetic material are directly used for calculation, the size is smaller than that of a conventional reservoir element using an electric signal or light. And low power consumption can be realized. One of the application fields of the information processing apparatus according to the present embodiment is the machine learning field in which demand is rapidly increasing in recent years. Therefore, its application fields are diverse. As an example, since stand-alone machine learning in a mobile device is possible, even if attention is paid only to this example, the technical and economic effects are great.

The embodiment disclosed this time should be considered as illustrative in all points and not restrictive. The scope of the present invention is defined by the terms of the claims, rather than the meanings described above, and is intended to include any modifications within the scope and meaning equivalent to the terms of the claims.

10 Input information cells (node selection unit, information writing unit)
20 Micro Magnetic Reservoir 21 Operation Node 30 Weighting Operation Element (Learning Unit)
40 Output information cell (information reading unit)

Claims

A magnetic reservoir, each including a plurality of nodes made of magnetic material, each node being magnetically coupled to at least one other node;
A node selection unit for selecting one or more nodes to which information is to be written from a plurality of nodes included in the magnetic substance reservoir;
An information writing unit for writing information to one or more selected nodes;
An information processing apparatus comprising: an information reading unit that reads information obtained by linear combination of each node included in the magnetic substance reservoir after writing the information.
The information processing apparatus according to claim 1, wherein the node selection unit randomly selects two or more nodes from a plurality of nodes included in the magnetic substance reservoir.
The learning unit according to claim 1 or 2, further comprising: a learning unit that learns linear weights in the linear combination so as to reproduce teacher information indicating an ideal output with respect to information written in the one or more nodes. Information processing device.
The magnetic body has one easy magnetization axis,
4. The information writing unit according to claim 1, wherein the information writing unit writes information by controlling a magnetization direction of a magnetic material that constitutes a node selected by the node selection unit. 5. Information processing device.
5. The magnetization direction of the magnetic material constituting each node is autonomously determined according to the magnetization direction of the magnetic material constituting another node included in the magnetic material reservoir after the information is written. Information processing device.
The information processing apparatus according to any one of claims 1 to 4, wherein the node is arranged on a lattice point of a rectangular lattice, a triangular lattice, a hexagonal lattice, or a rhombus lattice.
The information processing apparatus according to any one of claims 1 to 4, wherein the nodes are arranged aperiodically.
The magnetic reservoir includes a first layer including a node to which information is to be written, and a second layer including a plurality of nodes configured by a magnetic material whose magnetization direction is autonomously determined according to information written to the node. The information processing apparatus according to any one of claims 1 to 7.
The information processing apparatus according to claim 1, wherein the magnetic anisotropy of each node is periodically changed with respect to time according to a set rule.
Select one or more nodes to which information is to be written for a magnetic reservoir, each containing a plurality of nodes made of magnetic material, each node being magnetically coupled to at least one other node ,
Write information to one or more selected nodes,
An information processing method of reading information obtained by linear combination of each node included in the magnetic substance reservoir after writing the information.