Arithmetic device, common arithmetic setting device, common arithmetic setting method and program

Abstract

To enable storage capacity required for storing values defining functions to be relatively small.SOLUTION: An arithmetic device is provided, comprising a plurality of function arithmetic operation units performing arithmetic operations using functions, and a common function information storage unit storing values defining functions for each group of the grouped function arithmetic operation units.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to a computing device and a common computing setting device. , together The present invention relates to a calculation setting method and a program. [Background technology]

[0002] A technique has been proposed in which a nonlinear element is used for weighting signals at the edges of a neural network, and the parameter values ​​of the nonlinear element are the learning targets (see Non-Patent Document 1). [Prior art documents] [Non-patent literature]

[0003] [Non-Patent Document 1] Xiaolei Zhang, Yan Zhao, Kai Guo, Gaoliang Li and Nianmao Deng, "An Adaptive B-Spline Neural Network and Its Application in Terminal Sliding Mode Control for a Mobile Satcom Antenna Inertially Stabilized Platform", Sensors 2017, 17, 978,2017 Summary of the Invention [Problem to be solved by the invention]

[0004] In the case of nonlinear elements, the number of parameters is greater because the degree of freedom is higher than in the case of linear elements. When nonlinear functions are used in many places, such as when nonlinear functions using nonlinear elements are used to weight signals at edges of a neural network, it is preferable to make the memory capacity for storing parameters as small as possible.

[0005] One example of the object of the present invention is to provide a calculation device and a common calculation setting device that can store values ​​that define functions, such as parameters of nonlinear functions, in a relatively small memory capacity. , together The present invention provides a calculation setting method and a program. [Means for solving the problem]

[0006] According to a first aspect of the present invention, a calculation device includes a plurality of function calculation units that perform calculations using functions, and a function setting unit that sets the functions for each group into which the function calculation units are grouped. Part of a set of values a common function information storage unit for storing a value; an individual function information storage unit that stores, for each of the function calculation units, values ​​other than the values ​​stored in the common function information storage unit among a set of values ​​that define the function; Equipped with.

[0007] The function calculation unit performs calculations using the functions indicated by parameter values ​​set in a function template having parameters, and the common function information storage unit stores, for each group, parameter values ​​common to the function calculation units included in the group. , as the value that defines the function It may be arranged to store the information.

[0008] The common function information storage unit, for each group, is the value that defines the function The argument values ​​and function values ​​of the function The argument value and the function value are Linked Function Table At The function calculation unit may store the function table and perform calculations using the functions by referring to the function table.

[0010] According to a second aspect of the present invention, a common calculation setting device includes a parameter value individual model calculation device having a plurality of first function calculation units performing calculations using functions represented by parameter values ​​set in a function template having parameters, and an individual parameter value storage unit storing parameter values ​​for each of the first function calculation units, the parameter value individual model calculation device including: a grouping unit that groups the first function calculation units based on a similarity of the functions for each of the first function calculation units; a common parameter value calculation unit that calculates, for each group of the first function calculation units, a common parameter value that is a parameter value of one function to be used instead of the functions in all of the first function calculation units included in that group; and a parameter value common model setting unit that sets groups for the second function calculation units in accordance with the groups of the first function calculation units and stores the common parameter value in the common parameter value storage unit, the parameter value individual model calculation device including: a plurality of second function calculation units performing calculations using functions having parameters, and a common parameter value storage unit storing parameter values ​​for each group into which the second function calculation units are grouped.

[0012] The present invention 3 According to this aspect, the common calculation setting device is a parameter value common model calculation device having a plurality of function calculation units that perform calculations using functions having parameters, and a common parameter value storage unit that stores parameter values ​​for each group into which the function calculation units are grouped, the parameter value common model calculation device comprising: a group provisional setting unit that performs a plurality of groupings of the function calculation units; a learning control unit that causes the parameter value common model calculation device to learn the parameter values ​​for each group of the function calculation units for each of the groupings; a grouping determination unit that selects one of the groupings based on the learning results for each of the groupings; and a parameter value common model setting unit that sets groups according to the selected grouping for the function calculation units of the parameter value common model calculation device, and stores parameter values ​​in the learning results for the selected grouping in the common parameter value storage unit.

[0014] The present invention 4According to this aspect, the common calculation setting method includes: grouping the first function calculation units of a parameter value individual model calculation device, the first function calculation units being provided with a plurality of first function calculation units that perform calculations using functions represented by parameter values ​​set in a function template having parameters, and an individual parameter value storage unit that stores a parameter value for each of the first function calculation units, based on a similarity of the functions for each of the first function calculation units; calculating, for each group of the first function calculation units, a common parameter value that is a parameter value of one function that is used in place of the functions in all of the first function calculation units included in that group; and setting, for the second function calculation units of a parameter value common model calculation device, the second function calculation units being provided with a plurality of second function calculation units that perform calculations using functions having parameters, and a common parameter value storage unit that stores a parameter value for each of the groups into which the second function calculation units are grouped, a group corresponding to the group of the first function calculation units, and storing the common parameter value in the common parameter value storage unit.

[0016] The present invention 5 According to this aspect, the common calculation setting method includes: grouping the function calculation units in a parameter value common model calculation device having a plurality of function calculation units that perform calculations using functions having parameters, and a common parameter value storage unit that stores parameter values ​​for each group into which the function calculation units are grouped, into a plurality of groups; calculating, for each grouping, the parameter values ​​for each group of the function calculation units by learning; selecting one of the groupings based on the learning results for each grouping; setting groups according to the selected grouping for the function calculation units of the parameter value common model calculation device, and storing parameter values ​​in the learning results for the selected grouping in the common parameter value storage unit.

[0017] The present invention 6According to this aspect, the program causes a computer to execute the following steps: grouping the first function calculation units of a parameter value individual model calculation device, the first function calculation units being provided with a plurality of first function calculation units that perform calculations using functions represented by parameter values ​​set in a function template having parameters, and an individual parameter value storage unit that stores a parameter value for each of the first function calculation units, based on a similarity of the functions for each of the first function calculation units; calculating, for each group of the first function calculation units, a common parameter value that is a parameter value of one function that is used in place of the functions in all of the first function calculation units included in that group; and setting, for the second function calculation units of a parameter value common model calculation device, the second function calculation units being provided with a plurality of second function calculation units that perform calculations using functions having parameters, and a common parameter value storage unit that stores a parameter value for each of the groups into which the second function calculation units are grouped, a grouping according to the groups of the first function calculation units, and storing the common parameter value in the common parameter value storage unit.

[0019] The present invention 7 According to this aspect, the program causes a computer to execute the following steps of grouping the function calculation units of a parameter value common model calculation device having a plurality of function calculation units that perform calculations using functions having parameters, and a common parameter value storage unit that stores parameter values ​​for each group into which the function calculation units are grouped, grouping the function calculation units in a plurality of ways, calculating, for each of the groupings, the parameter values ​​for each group of the function calculation units by learning, selecting one of the groupings based on the learning results for each of the groupings, setting groups according to the selected grouping for the function calculation units of the parameter value common model calculation device, and storing parameter values ​​in the learning results for the selected grouping in the common parameter value storage unit. Effect of the Invention

[0020] The above-mentioned calculation device and common calculation setting device , togetherAccording to the calculation setting method and program, the memory capacity required to store values ​​that define the function can be made relatively small. [Brief description of the drawings]

[0021] [Figure 1] 1 is a diagram illustrating an example of a device configuration of an information processing system according to a first embodiment. [Diagram 2] FIG. 2 is a diagram illustrating an example of a configuration of a portion of a neural network according to the first embodiment. [Diagram 3] FIG. 2 is a diagram illustrating an example of the configuration of a parameter value individual model calculation device according to the first embodiment. [Figure 4] 1 is a diagram illustrating an example of the configuration of a parameter value commonization device according to a first embodiment. [Diagram 5] FIG. 2 is a diagram illustrating an example of the configuration of a parameter value common model calculation device according to the first embodiment. [Figure 6] 1 is a diagram illustrating an example of the configuration of a tabular device according to a first embodiment. [Figure 7] FIG. 4 is a diagram illustrating an example of a function indicated by a common parameter value according to the first embodiment. [Figure 8] FIG. 4 is a diagram illustrating an example of quantization of a function indicated by a common parameter value according to the first embodiment. [Figure 9] 4 is a diagram showing an example of a function table generated by a table calculation unit according to the first embodiment; FIG. [Figure 10] FIG. 1 is a diagram illustrating an example of the configuration of a table common model computing device according to a first embodiment. [Figure 11] 10 is a flowchart showing an example of a processing procedure when the information processing system according to the first embodiment performs setting of a parameter value common model calculation device and a table common model calculation device. [Figure 12] 5 is a flowchart showing an example of a processing procedure for standardizing a parameter value performed by the parameter value standardizing device according to the first embodiment. [Figure 13] 5 is a flowchart showing an example of a processing procedure for the tabulation device according to the first embodiment to convert a function into a function table. [Figure 14] 10 is a flowchart showing an example of a processing procedure of a common method weight processing unit during operation of the parameter value common model calculation device according to the first embodiment. [Figure 15] 10 is a flowchart showing an example of a processing procedure of a common table method weight processing unit during operation of the table common model computing device according to the first embodiment. [Figure 16] FIG. 1 is a diagram illustrating an example of the configuration of a parameter value mixed model calculation device according to a first embodiment. [Figure 17] FIG. 2 is a diagram illustrating an example of the configuration of a table mixed model calculation device according to the first embodiment. [Figure 18] FIG. 2 is a diagram illustrating an example of the configuration of a parameter table-combined model calculation device according to the first embodiment. [Figure 19] FIG. 1 is a diagram illustrating an example of a configuration of a neural network system according to a first embodiment. [Figure 20] FIG. 11 is a diagram illustrating an example of a device configuration of an information processing system according to a second embodiment. [Figure 21] FIG. 11 is a diagram illustrating an example of the configuration of a group setting method learning control device according to a second embodiment. [Figure 22] 13 is a flowchart showing an example of a processing procedure when the information processing system according to the second embodiment performs setting of a parameter value common model calculation device and a table common model calculation device. [Figure 23] 13 is a flowchart showing an example of a processing procedure in which the group setting method learning control device according to the second embodiment controls the parameter value common model calculation device to perform learning. [Figure 24] FIG. 1 is a schematic block diagram illustrating a configuration of a computer according to at least one embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0022] The following describes embodiments of the present invention, but the following embodiments do not limit the scope of the invention. Furthermore, not all of the combinations of features described in the embodiments are necessarily essential to the solution of the invention.

[0023] First Embodiment Fig. 1 is a diagram showing an example of a device configuration of an information processing system according to the first embodiment. In the configuration shown in Fig. 1, the information processing system 1 includes a parameter value individual model calculation device 11, a parameter value individual learning control device 12, a parameter value commonization device 13, a parameter value common model calculation device 14, a parameter value common learning control device 15, a tabulation device 16, and a table common model calculation device 17.

[0024] The information processing system 1 is a system for setting functions of a computing device that performs a calculation using a deformable function. Among the components in Fig. 1, the parameter value individual model computing device 11, the parameter value common model computing device 14, and the table common model computing device 17 correspond to examples of computing devices that perform a calculation using a deformable function. A computing device that performs a calculation using a deformable function is also simply referred to as a computing device. In the following, a case will be described in which the arithmetic device handled by the information processing system 1 is a device that implements a neural network that weights signals at edges using a transformable function. However, the arithmetic device handled by the information processing system 1 is not limited to this.

[0025] Fig. 2 is a diagram showing an example of a partial configuration of a neural network. In Fig. 2, two hidden layer nodes and edges between the nodes of a hierarchical and fully connected neural network are shown. However, when a neural network is implemented in a calculation device handled by the information processing system 1, the type of the implemented neural network is not limited to a specific type.

[0026] In the configuration shown in FIG. 2, the neural network system 110 includes a node operation unit 111 and a weight processing unit 112. The node operation unit 111 functions as a node of a neural network. In the following, an example will be described in which the node operation unit 111 inputs the sum of the inputs to the node operation unit 111 itself to an activation function and outputs the obtained function value. However, when a neural network is implemented in a calculation device handled by the information processing system 1, the type of node of the implemented neural network is not limited to a specific type.

[0027] The weight processing unit 112 functions as an edge of the neural network. In the following, an example will be described in which the weight processing unit 112 weights signals (numerical values ​​transmitted from node to node) at the edges. In particular, an example will be described in which the weight processing unit 112 performs weighting calculations using a transformable function. However, when a neural network is implemented in a calculation device handled by the information processing system 1, the method of weighting signals between nodes is not limited to a specific method.

[0028] Fig. 3 is a diagram showing an example of the configuration of the parameter value individual model calculation device 11. In the configuration shown in Fig. 3, the parameter value individual model calculation device 11 includes a node calculation unit 111 and an individual method weight processing unit 120. The individual method weight processing unit 120 includes a function calculation unit 121 and an individual parameter value storage unit 122. The parameter value individual model calculation device 11 corresponds to an example of a neural network system 110. The individual scheme weight processing unit 120 corresponds to an example of a weight processing unit 112.

[0029] The individual scheme weighting processor 120 weights signals at edges using a transformable function. Specifically, a function calculator 121 calculates a function having parameters. An individual parameter value storage unit 122 stores parameter values ​​of the function. The function calculated by the function calculator 121 is specified by setting the parameter values ​​stored in the individual parameter value storage unit 122 to the parameters of the function. Specifying a function is also referred to as determining the function.

[0030] The function calculation unit 121 inputs the value output from the node calculation unit 111 to a function specified by the parameter value, and outputs the calculated function value to the node calculation unit 111. The calculation performed by the function calculation unit 121 functions as weighting at the edges of the neural network, and the parameter values ​​stored in the individual parameter value storage unit 122 are updated by learning. The function calculation unit 121 of the parameter value individual model calculation device 11 corresponds to an example of a first calculation unit.

[0031] A function having parameters is also called a function template. A function is determined by setting parameter values ​​to the parameters of the function template. The parameter values ​​of the functions stored in the individual parameter value storage unit 122 can be said to be parameter values ​​that specify the function by being set to the parameters of the function template. All of the function calculation units 121 may have one function template with the same content, but the function templates provided by the function calculation units 121 may differ from one another. A function determined by setting the parameter values ​​stored in the individual parameter value storage unit 122 of the individual scheme weighting unit 120 to the parameters of the function template is also referred to as the function of that individual scheme weighting unit 120.

[0032] The function that is calculated by the function calculation unit 121 may be a nonlinear function or a piecewise linear function. For example, the function that is calculated by the function calculation unit 121 may be a B-spline function. The B-spline function here is a function whose graph shows a B-spline curve. Since the B-spline function is a piecewise polynomial, updating one parameter does not change the shape of the entire function, but only changes the function value in a part of the domain. In this respect, when the B-spline function is used as the subject of learning updates, it is expected that learning will converge relatively quickly. However, the function that the function calculation unit 121 performs the calculation is not limited to the B-spline function.

[0033] By using a relatively complex function such as a nonlinear function or a piecewise linear function as an update target in learning, as in the parameter value individual model calculation device 11, more information can be expressed by one node compared to the case of weighting by multiplication of a weighting coefficient. In this respect, it is expected that by using a relatively complex function as an update target in learning, the number of nodes in the neural network can be made relatively small.

[0034] On the other hand, when a relatively complicated function is to be updated in learning, the number of parameters of the function increases, and the memory capacity for storing the parameter values ​​becomes large. From the viewpoint of miniaturization of the device or in the case of implementation with relatively low performance, it is preferable to make the memory capacity as small as possible.

[0035] Therefore, the information processing system 1 calculates a parameter value for each group of edges (weight processing unit 112 in FIG. 2) from the parameter values ​​of the function of the learned parameter value individual model calculation device 11, and sets the parameter value in the parameter value common device 13. The parameter value individual learning control device 12 causes the parameter value individual model calculation device 11 to perform learning. As a learning method for the parameter value individual model calculation device 11, a known method such as backpropagation can be used.

[0036] The parameter value commonization device 13 groups the individual scheme weighting processing units 120 based on the parameter values ​​stored in the individual parameter value storage unit 122 of the parameter value individual model calculation device 11, and calculates a parameter value for each group. The parameter value commonization device 13 sets the calculated parameter values ​​for each group in the parameter value common model calculation device 14. The parameter value commonization device 13 corresponds to an example of a common calculation setting device.

[0037] Fig. 4 is a diagram showing an example of the configuration of the parameter value commonization device 13. In the configuration shown in Fig. 4, the parameter value commonization device 13 includes a grouping unit 311, a common parameter value calculation unit 312, and a parameter value common model setting unit 313. The grouping unit 311 groups the individual scheme weighting units 120 based on the parameter values ​​stored in the individual parameter value storage unit 122 of the parameter value individual model calculation device 11 .

[0038] Each of the individual scheme weight processors 120 includes one function calculator 121. Therefore, the grouping of the individual scheme weight processors 120 also corresponds to the grouping of the function calculators 121 (first calculators) of the individual scheme weight processors 120. Furthermore, one function is specified for one individual scheme weight processor 120 by the parameter values ​​stored in the individual parameter value storage section 122. Therefore, grouping of the individual scheme weight processors 120 also corresponds to grouping of the functions. Grouping is also called grouping.

[0039] The grouping unit 311 calculates an index value of the similarity between functions of the individual scheme weighting unit 120. Then, the grouping unit 311 determines groups of the individual scheme weighting unit 120 such that the individual scheme weighting unit 120 having a similarity higher than a predetermined condition is included in the same group.

[0040] The method by which the grouping unit 311 calculates the index value of the similarity of the functions is not limited to a specific method. For example, the grouping unit 311 may sample function values ​​for each of a plurality of functions, and calculate the magnitude of the difference between the function values ​​of two functions as an index value that becomes smaller as the functions become more similar.

[0041] The grouping unit 311 may calculate a function value for a predetermined input value to the function and use the calculated function value as a sampling value. The grouping unit 311 may sample the function value for each of a plurality of input values ​​to the function to calculate an index value. Alternatively, in the case where the functions can be evaluated as being more similar as the difference in parameter values ​​is smaller, the grouping unit 311 may calculate the magnitude of the difference in parameter values ​​as an index value that becomes smaller as the functions become more similar.

[0042] The method by which the grouping unit 311 groups the individual scheme weight processing units 120 is not limited to a specific method. For example, the grouping unit 311 may determine groups of the individual scheme weighting processors 120 so that, for any combination of two individual scheme weighting processors 120 included in the same group, the index value, the smaller the value of which is the more similar the functions, is equal to or smaller than a predetermined threshold value.

[0043] The common parameter value calculation unit 312 calculates, for each group of individual scheme weight processing units 120 set by the grouping unit 311, a parameter value common to the individual scheme weight processing units 120 included in that group. A parameter value common to the individual scheme weight processing units 120 included in a group is also referred to as a common parameter value. It can be said that the common parameter value is a parameter value of one function used in place of functions in all function calculation units 121 included in the group of function calculation units 121 of the individual scheme weight processing unit 120.

[0044] The method by which the common parameter value calculation section 312 calculates the parameter values ​​is not limited to a specific method. The common parameter value calculation unit 312 may calculate a parameter value of a function that approximates the functions of all the individual scheme weighting processors 120 included in one group. For example, the common parameter value calculation unit 312 may calculate a parameter value of an average function of the functions of all the individual scheme weighting processors 120 included in one group. The function of an individual scheme weighting processor 120 here is a function specified by a parameter value stored in the individual parameter value storage unit 122 of that individual scheme weighting processor 120.

[0045] Further, for example, the common parameter value calculation unit 312 may sample the functions of all the individual scheme weighting processors 120 included in one group for each of a plurality of input values ​​to the function and calculate an average for each input value.The common parameter value calculation unit 312 may then calculate, as an average function, a function that indicates a curve that approximates points (coordinates) obtained for the number of input values.

[0046] The parameter value common model setting section 313 sets the parameter values ​​for each group calculated by the common parameter value calculation section 312 in the parameter value common model calculation device 14 . In addition, the parameter value common model setting unit 313 determines a group of common scheme weight processing units, which are weight processing units in the parameter value common model calculation device 14, based on the group of the individual scheme weight processing units 120, and sets the group to each of the common scheme weight processing units.

[0047] For example, the parameter value common model computing device 14 may be configured so that the nodes and edges of the neural network are in one-to-one correspondence with the nodes and edges of the neural network in the parameter value individual model computing device 11. The parameter value common model setting section 313 may then determine a group of common method weighting processing sections in the parameter value common model computing device 14 so that the sections included in the group set in the parameter value individual model computing device 11 and the sections included in the group set in the parameter value common model computing device 14 are at the same level as the nodes and edges of the neural network.

[0048] Fig. 5 is a diagram showing an example of the configuration of the parameter value common model calculation device 14. In the configuration shown in Fig. 5, the parameter value common model calculation device 14 includes a node calculation unit 111, a common parameter value storage unit 130, and a common method weight processing unit 140. The common method weight processing unit 140 includes a function calculation unit 121, a group ID storage unit 141, and a parameter value acquisition unit 142.

[0049] The parameter value common model calculation device 14 corresponds to an example of a neural network system 110. The common method weight processing unit 140 corresponds to an example of a weight processing unit 112. Except for the counterpart from which the function calculation unit 121 acquires the parameter values, the function calculation unit 121 is similar to the function calculation unit 121 of the parameter value individual model calculation device 11. In the parameter value common model calculation device 14, the function calculation unit 121 acquires the parameter values ​​from the common parameter value storage unit 130 via the parameter value acquisition unit 142. The function calculation unit 121 of the parameter value common model calculation device 14 corresponds to an example of a second calculation unit. Each common scheme weight processor 140 includes one function calculator 121. Therefore, the grouping of the common scheme weight processors 140 also corresponds to the grouping of the function calculators 121 (second calculators) of the common scheme weight processors 140.

[0050] The common parameter value storage unit 130 stores the common parameter value. That is, the common parameter value storage unit 130 stores the parameter value calculated by the common parameter value calculation unit 312 of the parameter value commonization device 13 for each group (in the example of FIG. 5, the group of the common scheme weight processing unit 140). The parameter values ​​for each group stored in the common parameter value storage unit 130 are an example of values ​​that define a function. The common parameter value storage unit 130 is an example of a common function information storage unit.

[0051] The common parameter value storage unit 130 stores the common parameter value in association with a group ID. The group ID is identification information for identifying a group. The common parameter value storage unit 130 stores, for each group of the common scheme weight processing unit 140, a group ID indicating the group and the common parameter value of the group.

[0052] The method in which the common parameter value storage unit 130 associates the common parameter values ​​with the group IDs and stores them is not limited to a specific method. For example, the common parameter value storage unit 130 may store data in a table format including a column for group IDs and a column for common parameter values, in which the group IDs are clearly indicated. Alternatively, the groups in the common scheme weight processing unit 140 are ordered, and the common parameter value storage unit 130 may store the common parameter values ​​in an order according to the order of the groups, in which the group IDs are not clearly indicated.

[0053] The common scheme weighting unit 140 weights the signals at the edges using a transformable function, similar to the individual scheme weighting unit 120 of the parameter value individual model calculation unit 11 . However, in the common scheme weight processor 140, the parameter values ​​of the function template are read out from a different source than in the individual scheme weight processor 120. As described above, the individual scheme weight processor 120 reads out parameter values ​​from its own individual parameter value storage unit 122, and sets them as parameters of the function template. On the other hand, the common scheme weight processor 140 reads out parameter values ​​of the group to which the common scheme weight processor 140 itself belongs, from among the parameter values ​​stored for each group in the common parameter value storage unit 130, and sets them as parameters of the function template.

[0054] In other respects, the common scheme weight processor 140 is similar to the individual scheme weight processor 120 . A function determined by setting the parameter values ​​read by the common scheme weight processor 140 from the common parameter value storage unit 130 as parameters of a function template is also referred to as the function of the common scheme weight processor 140 .

[0055] 5, the group ID storage unit 141 stores a group ID for identifying a group of the common scheme weight processing unit 140 set by the parameter value common model setting unit 313 of the parameter value commonization device 13.

[0056] The parameter value acquiring unit 142 reads out the group ID stored in the group ID storage unit 141. Then, the parameter value acquiring unit 142 reads out the common parameter values ​​associated with the group ID read out from the group ID storage unit 141, among the common parameter values ​​stored in the common parameter value storage unit 130. In this way, the common scheme weight processing unit 140 acquires the common parameter values ​​of the group to which the common scheme weight processing unit 140 itself belongs.

[0057] The function operation unit 121 specifies a function by setting the common parameter value acquired by the parameter value acquisition unit 142 as the parameter of the function template. Then, the function operation unit 121 inputs the value from the node operation unit 111 to the specified function to calculate a function value. The function operation unit 121 outputs the calculated function value to the node operation unit 111.

[0058] In the parameter value common model calculation device 14, the parameter values ​​of the function template are shared for each group of the common method weight processing unit 140, so that a smaller memory capacity is required for storing the parameter values ​​than in the parameter value individual model calculation device 11. As in the case of the parameter value individual model calculation device 11, in the parameter value common model calculation device 14, all the function calculation units 121 may each have one function template with the same content, but the function templates provided by the function calculation units 121 may be different. As long as the same parameter values ​​can be set to the parameters of the function templates of the function calculation units 121 included in the same group, the functions specified by the setting of the parameter values ​​may be different for each function calculation unit 121.

[0059] After the parameter value commonization device 13 sets the parameter value common model calculation device 14, the parameter value common learning control device 15 may cause the parameter value common model calculation device 14 to perform learning to improve the accuracy of the processing of the parameter value common model calculation device 14. For example, when the parameter value common model calculation device 14 performs class classification, the processing accuracy referred to here is classification accuracy (for example, accuracy rate). However, the processing performed by the parameter value common model calculation device 14 is not limited to a specific processing.

[0060] The parameter value common model calculation device 14 may use a learning method similar to that used by the parameter value individual model calculation device 11. Then, instead of updating the parameter values ​​stored in the individual parameter value storage unit 122, the parameter values ​​stored in the common parameter value storage unit 130 may be updated.

[0061] The parameter value common model calculation device 14 may be operated by, for example, mounting the parameter value common model calculation device 14 in a product. Alternatively, the processing load of the function calculation performed by the function calculation unit 121 may be reduced by making the function calculation into a function table. The function table here is data in which argument values ​​of a function (input values ​​to a function) and function values ​​for the argument values ​​(output values ​​of the function) are linked.

[0062] 6 is a diagram showing an example of the configuration of the tabulation device 16. In the configuration shown in FIG. The tabulation device 16 generates a function table of functions indicated by the common parameter values ​​stored in the common parameter value storage unit 130 of the parameter value common model calculation device 14. The function indicated by the parameter values ​​here is a function obtained by setting the parameter values ​​to the parameters of a function template. The tabulation device 16 then sets the generated function table in the table common model calculation device 17. The table common model calculation device 17 is configured as a device in which the calculation of a function obtained by setting a common parameter value to a parameter of a function template in the parameter value common model calculation device 14 is replaced with the calculation of a function by referring to a function table.

[0063] The table calculation unit 321 generates a function table of functions indicated by the common parameter values ​​for each common parameter value stored in the common parameter value storage unit 130 of the parameter value common model calculation device 14. Therefore, the table calculation unit 321 generates a function table indicating functions for weight calculation for each group of the common method weight processing unit 140 in the parameter value common model calculation device 14.

[0064] Fig. 7 is a diagram showing an example of a function indicated by a common parameter value. The horizontal axis of the graph in Fig. 7 indicates an argument value (an input value to the function). The vertical axis indicates a function value (an output value of the function). The argument value and the function value may be expressed as a decimal number such as a floating point number.

[0065] Fig. 8 is a diagram showing an example of quantization of a function indicated by a common parameter value. The quantization referred to here means approximately expressing an analog value with a discrete digital value. The horizontal axis of the graph in Fig. 8 indicates the argument value, and the vertical axis indicates the function value. In the example in Figure 8, the domain of the function is 2 N The input data is divided into intervals. The argument value in each interval is quantized to be represented as an N-bit bit string, such as (0 000)2, (0 001)2, (0 010)2, . . ., (1 111)2, where N is a positive integer. For example, the argument value may be quantized to be represented as an N-bit fixed-point number.

[0066] By generating a function table of quantized functions in this manner, the search key for referencing the function table is represented by an N-bit bit string, making it relatively easy to search the function table. In addition, on the horizontal axis of Figure 8, (0) 10 =(0···000)2, (1) 10 =(0 . . . 001)2, (2) 10 The quantized argument values ​​are shown in decimal, such as =(0···010)2, ···.

[0067] In the example of Figure 8, the range of the function is also 2 N The function is divided into intervals and quantized so that the function value in each interval is represented by an N-bit integer. This quantization allows the function value obtained by referencing the function table (the output value from the function table) to be represented by an N-bit bit string, making it relatively easy to communicate the search results of the function table.

[0068] Fig. 9 is a diagram showing an example of a function table generated by table calculation unit 321. In the example of Fig. 9, the function table is represented as tabular data, and each row is associated with an argument value and a function value for that argument value. The argument value and the function value are each quantized to an N-bit integer as in the example of Fig. 8.

[0069] The table common model setting unit 322 sets the function table for each group, which is generated by the table calculation unit 321, in the table common model calculation device 17. In addition, the table common model setting unit 322 determines a group of common table method weight processing units, which are weight processing units in the table common model calculation device 17, based on the group of the common method weight processing units 140, and sets the group to each of the common table method weight processing units.

[0070] For example, the table common model computing device 17 may be configured so that the nodes and edges of the neural network are in one-to-one correspondence with the nodes and edges of the neural network in the parameter value common model computing device 14. The table common model setting unit 322 may determine the group of the common table method weight processing unit of the table common model computing device 17 so that each unit included in the group set in the parameter value common model computing device 14 and each unit included in the group set in the table common model computing device 17 are at the same level of the nodes and edges of the neural network. The table common model calculation device 17 in which the function table is set can retrieve the function value in the function table using the argument value of the function as a search key, thereby reading out the function value for that argument value.

[0071] Fig. 10 is a diagram showing an example of the configuration of the table common model calculation device 17. In the configuration shown in Fig. 10, the table common model calculation device 17 includes a node calculation unit 111, a common table storage unit 150, and a common table method weight processing unit 160. The common table method weight processing unit 160 includes a table reference unit 161 and a group ID storage unit 141.

[0072] The table common model calculation device 17 corresponds to an example of the neural network system 110. The common table method weight processing unit 160 corresponds to an example of the weight processing unit 112. Each common table method weight processor 160 includes one table reference unit 161. Therefore, grouping of common table method weight processors 160 also corresponds to grouping of table reference units 161.

[0073] The common table storage unit 150 stores a function table for each group of common table method weight processors 160 set by the table common model setting unit 322 of the tabulation device 16. The function table stored by the common table storage unit 150 for each group of common table method weight processors 160 is used as a function table common to the common table method weight processors 160 included in that group. A function table common to the common table method weight processors 160 included in a group is also referred to as a common table. The argument values ​​and function values ​​in the function table stored in the common table storage unit 150 are examples of values ​​that define a function. The common table storage unit 150 is an example of a common function information storage unit.

[0074] The common table storage unit 150 stores the common table in association with a group ID. For each group of the common table method weight processing unit 160, the common table storage unit 150 stores a group ID indicating the group and the common table of the group.

[0075] The method in which the common table storage unit 150 stores the common table in association with the group ID is not limited to a specific method. For example, the common table storage unit 150 may store data in a tabular format including a column of group IDs and a column of a common table, in which the group ID is clearly indicated. Alternatively, the groups in the common table method weight processing unit 160 are ordered, and the common table storage unit 150 may store data in which the group ID is not clearly indicated, in which case the groups are ordered.

[0076] The common table weighting processor 160 weights signals at edges using a transformable function, similar to the common table weighting processor 140 of the parameter value common model calculation device 14 . However, the common table method weight processing unit 160 uses a value from the node calculation unit 111 as an argument value and acquires a function value for that argument value in a manner different from that of the common method weight processing unit 140. The common method weight processing unit 140 specifies a function by setting the common parameter value read out from the common parameter value storage unit 130 as a parameter of a function template, and inputs the argument value into the specified function to calculate a function value. On the other hand, the common table method weight processing unit 160 refers to the common table stored in the common table storage unit 150 and reads out a function value linked to the argument value.

[0077] In other respects, the common table weight processor 160 is similar to the common table weight processor 140 . The function indicated by the common table referred to by the common table method weight processor 160 is also referred to as the function of the common table method weight processor 160 .

[0078] 10, the group ID storage unit 141 stores a group ID for identifying a group of the common table method weight processing unit 160, which is set by the table common model setting unit 322 of the tabulation device 16.

[0079] The table reference unit 161 reads out the group ID stored in the group ID storage unit 141. Then, the table reference unit 161 refers to a common table associated with the group ID read out from the group ID storage unit 141, among the common tables stored in the common table storage unit 150. The table reference unit 161 quantizes a value from the node operation unit 111, which is an argument value of the function, into an N-bit bit string, and searches for a function value in the common table using the quantized bit string as a search key. Through this search, the table reference unit 161 reads out a function value in the case of the argument value indicated by the search key, in the form of the quantized bit string. The table reference unit 161 converts the function value indicated by the quantized bit string into a function value in the case where no quantization is performed. The table reference unit 161 outputs the obtained function value to the node operation unit 111.

[0080] In the table common model calculation device 17, since function values ​​are read by referring to a function table, it is expected that the processing load for obtaining function values ​​is relatively light even when the function is complex, and the processing time for obtaining the function is relatively short. Moreover, a relatively large storage capacity is required to represent the function in the form of a function table. In this regard, in the table common model calculation device 17, the common table storage unit 150 only needs to store one common table for each group of the common table method weight processing units 160. Therefore, in the table common model calculation device 17, even if the number of common table method weight processing units 160 is large, the number of common tables stored in the common table storage unit 150 is relatively small, and it is expected that the storage capacity for storing the common tables will be relatively small. The table common model calculation device 17 may be operated by, for example, mounting the table common model calculation device 17 on a product.

[0081] FIG. 11 is a flowchart showing an example of a processing procedure when the information processing system 1 performs setting of the parameter value common model calculation device 14 and the table common model calculation device 17. 11, the parameter value individual model calculation device 11 performs learning under the control of the parameter value individual learning control device 12 (step S111). The parameter value individual model calculation device 11 updates the parameter values ​​stored in the individual parameter value storage unit 122 through learning. This learning corresponds to learning of a neural network implemented in the parameter value individual model calculation device 11. In the process of step S111, it is sufficient that the parameter value individual model calculation device 11 and the parameter value individual learning control device 12 are present among the devices shown in FIG. 1, and there is no need to prepare any other devices.

[0082] Next, the parameter value commonization device 13 commonizes the parameter values ​​stored in the individual parameter value storage unit 122 of the parameter value individual model calculation device 11, and sets the commonization result in the parameter value common model calculation device 14 (step S112). The parameter value commonization device 13 sets the commonized parameter values ​​and the common method weight processing unit 140 group in the commonization of the parameter values ​​in the parameter value common model calculation device 14. In the processing of step S112, it is sufficient that the parameter value individual model calculation device 11, the parameter value commonization device 13, and the parameter value common model calculation device 14 are present among the devices shown in FIG. 1, and there is no need to prepare any other devices.

[0083] Next, the parameter value common model calculation device 14 performs learning under the control of the parameter value common learning control device 15 (step S113). The parameter value common model calculation device 14 updates the common parameter values ​​stored in the common parameter value storage unit 130 through learning. This learning corresponds to learning of a neural network implemented in the parameter value common model calculation device 14. In the process of step S113, it is sufficient that the parameter value common model calculation device 14 and the parameter value common learning control device 15 are present among the devices shown in FIG. 1, and there is no need to prepare any other devices.

[0084] The parameter value common model calculation device 14 can be said to have already been trained in that it can be said to have transplanted the settings of the trained parameter value individual model calculation device 11. The learning of the parameter value common model calculation device 14 in step S113 corresponds to fine tuning that further improves the accuracy of the processing of the parameter value common model calculation device 14.

[0085] Next, the tabulation device 16 creates a function table of the functions indicated by the common parameter values ​​stored in the common parameter value storage unit 130 of the parameter value common model calculation device 14, and sets the function tabulation result in the table common model calculation device 17 (step S114). The tabulation device 16 sets the obtained function table and a group of the common table method weight processing unit 160 for each function table in the table common model calculation device 17. In the process of step S114, it is sufficient that the parameter value common model calculation device 14, the tabulation device 16, and the table common model calculation device 17 are present among the devices shown in FIG. 1, and there is no need to prepare any other devices. After step S14, the information processing system 1 ends the process of FIG.

[0086] 12 is a flowchart showing an example of a processing procedure for standardizing parameter values ​​by the parameter value standardization device 13. The parameter value standardization device 13 performs the processing of FIG. 12, the grouping unit 311 groups functions based on the similarity of the functions of the individual scheme weighting unit 120 of the parameter value individual model calculation device 11 (step S121). Since one function is shown for one individual scheme weighting unit 120, the individual scheme weighting unit 120 is also grouped by the grouping of functions performed by the grouping unit 311.

[0087] Next, the parameter value standardization device 13 starts a loop L11 in which processing is performed for each group of the individual scheme weight processor 120 (step S122). In the processing of the loop L11, the group of the individual scheme weight processor 120 that is the processing target is also referred to as a target group. In the process of loop L11, the common parameter value calculation unit 312 calculates the common parameter value of the target group (step S123). As described above, the common parameter value calculation unit 312 may calculate, as the common parameter value, the parameter value of a function that approximates the functions of all the individual scheme weighting units 120 included in the target group.

[0088] The parameter value common model setting unit 313 associates the common parameter values ​​calculated by the common parameter value calculation unit 312 with the group ID and stores them in the common parameter value storage unit 130 of the parameter value common model calculation device 14 (step S124). In addition, the parameter value common model setting unit 313 determines a group of common scheme weight processing units 140 corresponding to the target group, and stores the group ID in the group ID storage unit 141 of the common scheme weight processing units 140 included in the determined group (step S125).

[0089] As described above, the parameter value common model calculation device 14 may be configured so that the nodes and edges of the neural network are in one-to-one correspondence with the nodes and edges of the neural network in the parameter value individual model calculation device 11. Then, the parameter value common model setting unit 313 may determine a group of all the common scheme weight processing units 140 associated with the individual scheme weight processing units 120 included in the target group as the group of the common scheme weight processing units 140 corresponding to the target group.

[0090] Next, the parameter value common device 13 performs a termination process of the loop L11 (step S126). Specifically, the parameter value common device 13 judges whether or not the process of the loop L11 has been performed for all the groups obtained in step S121. If it is judged that there is a group for which the process of the loop L11 has not been performed, the process returns to step S123, and the parameter value common device 13 continues to perform the process of the loop L11 for the unprocessed group. On the other hand, if it is judged that the process of the loop L11 has been performed for all the groups, the parameter value common device 13 ends the loop L11. When the loop L11 is ended in step S126, the parameter value standardization device 13 ends the processing of FIG.

[0091] 13 is a flowchart showing an example of a processing procedure for converting a function into a function table by the tabulation device 16. The tabulation device 16 performs the processing of FIG. 13, the tabulation device 16 starts a loop L21 for performing processing for each group of the common scheme weight processor 140 (step S131). In the loop L21, the common scheme weight processor 140 group that is the processing target is also referred to as a target group.

[0092] In the processing of loop L21, the table calculation unit 321 creates a function table of functions indicated by common parameter values ​​associated with the group ID of the target group, among the common parameter values ​​stored in the common parameter value storage unit 130 (step S132). The table common model setting unit 322 associates the function table generated by the table calculation unit 321 with the group ID and stores it in the common table storage unit 150 of the table common model calculation device 17 (step S133). In addition, the table common model setting unit 322 determines a group of common table method weight processors 160 corresponding to the target group, and stores the group IDs in the group ID storage unit 141 of the common table method weight processors 160 included in the determined group (step S134).

[0093] As described above, the table common model calculation device 17 may be configured so that the nodes and edges of the neural network are in one-to-one correspondence with the nodes and edges of the neural network in the parameter value common model calculation device 14. Then, the table common model setting unit 322 may determine a group of all the common table method weight processing units 160 associated with the common method weight processing units 140 included in the target group as the group of the common table method weight processing units 160 corresponding to the target group.

[0094] Next, the tabulating device 16 performs a termination process of the loop L21 (step S135). Specifically, the tabulating device 16 judges whether or not the process of the loop L21 has been performed for all groups of the common scheme weight processing unit 140. If it is judged that there is a group for which the process of the loop L21 has not been performed, the process returns to step S132, and the tabulating device 16 continues to perform the process of the loop L21 for the unprocessed group. On the other hand, if it is judged that the process of the loop L21 has been performed for all groups, the tabulating device 16 ends the loop L21. When the loop L21 is ended in step S136, the tabulation device 16 ends the processing of FIG.

[0095] 14 is a flowchart showing an example of a processing procedure of the common method weight processing unit 140 during operation of the parameter value common model calculation device 14. The common method weight processing unit 140, which receives an input of a value from the node calculation unit 111, performs the process of FIG.

[0096] In the process of FIG. 14, the parameter value acquisition unit 142 reads out a group ID from the group ID storage unit 141 (step S141). Then, the parameter value acquisition unit 142 reads out, from the common parameter value storage unit 130, the common parameter value associated with the group ID read out from the group ID storage unit 141 (step S142).

[0097] The function calculation unit 121 sets the common parameter values ​​acquired by the parameter value acquisition unit to the parameters of the function template (step S143). By setting the parameter values ​​to the parameters of the function template, the function is specified. Then, the function operation unit 121 inputs the value from the node operation unit 111 to the specified function to calculate a function value (step S144). The function operation unit 121 outputs the calculated function value to the node operation unit 111. After step S144, the parameter value common model calculation device 14 ends the process of FIG.

[0098] 15 is a flowchart showing an example of a processing procedure of the common table method weight processing unit 160 during operation of the table common model calculation device 17. Upon receiving an input of a value from the node calculation unit 111, the common table method weight processing unit 160 performs the process of FIG.

[0099] In the process of FIG. 15, table reference unit 161 reads out a group ID from group ID storage unit 141 (step S151). Then, the table reference unit 161 converts the value from the node operation unit 111 by referring to the common table associated with the group ID read from the group ID storage unit 141 (step S152).

[0100] Specifically, the table reference unit 161 identifies a common table linked to the group ID read from the group ID storage unit 141, among the common tables stored in the common table storage unit 150. Then, the table reference unit 161 searches for a function value in the identified common table using the value from the node operation unit 111 as a search key, and reads out the function value linked to the value from the node operation unit 111. In this way, the table reference unit 161 obtains the function value when the value from the node operation unit 111 is input to the function as an argument value. The table reference unit 161 outputs the obtained function value to the node operation unit 111.

[0101] 7 to 9, when the values ​​stored in the common table are quantized, table reference unit 161 quantizes the values ​​from node operation unit 111 and searches for function values ​​in the common table using the quantized values ​​as a search key. Table reference unit 161 also performs an inverse operation of quantization on the function values ​​read from the common table and outputs the function values ​​after the inverse operation to node operation unit 111. Alternatively, the quantized values ​​may be used in the entire table common model calculation device 17. In this case, each common table method weight processing unit 160 does not need to perform quantization and inverse quantization. After step S152, the common table method weight processor 160 ends the process of FIG.

[0102] Instead of the parameter value common model calculation device 14, the information processing system 1 may include a device in which parameter values ​​are commonized for only some of the parameters of the function templates. 16 is a diagram showing an example of the configuration of the parameter value mixed model calculation device 24. The parameter value mixed model calculation device 24 corresponds to an example of a device in which parameter values ​​are standardized for only some of the parameters of the function template.

[0103] 16, parameter value mixed model calculation device 24 includes node calculation section 111, common parameter value storage section 210, and mixed method weight processing section 220. Mixed method weight processing section 220 includes function calculation section 221, group ID storage section 141, parameter value acquisition section 142, and individual parameter value storage section 222. The parameter value mixture model calculation device 24 corresponds to an example of the neural network system 110. The mixture method weight processing unit 220 corresponds to an example of the weight processing unit 112.

[0104] The common parameter value storage unit 210 stores parameter values ​​of function templates. Comparing the common parameter value storage unit 210 with the common parameter value storage unit 130 of the parameter value common model calculation device 14, the common parameter value storage unit 130 stores parameter values ​​for all parameters of a function template, whereas the common parameter value storage unit 210 stores parameter values ​​for only some of the parameters of a function template. In other respects, the common parameter value storage unit 210 is similar to the common parameter value storage unit 130. As in the case of the common parameter value storage unit 130, the parameter values ​​stored in the common parameter value storage unit 210 are also referred to as common parameter values.

[0105] The group ID storage unit 141 of the parameter value mixed model calculation device 24 stores a group ID, similar to the group ID storage unit 141 of the parameter value common model calculation device 14. In the parameter value mixed model calculation device 24, the group ID identifies a group of the mixing scheme weighting processing unit 220.

[0106] The parameter value acquisition unit 142 reads out from the common parameter value storage unit 210 the common parameter value associated with the group ID stored in the group ID storage unit 141 . The process performed by the parameter value acquisition section 142 of the parameter value mixed model calculation device 24 is similar to that of the parameter value acquisition section 142 of the parameter value common model calculation device 14, except that the parameter values ​​of only some of the parameters of the function template are read out.

[0107] The individual parameter value storage unit 222 stores parameter values ​​of the function template. Specifically, the individual parameter value storage unit 222 stores parameter values ​​of the function template other than the parameters whose parameter values ​​are stored in the common parameter value storage unit.

[0108] Comparing the individual parameter value storage unit 222 with the individual parameter value storage unit 122 of the parameter value individual model calculation device 11, the individual parameter value storage unit 122 stores parameter values ​​for all parameters of a function template, whereas the individual parameter value storage unit 222 stores parameter values ​​for only some of the parameters of a function template. In other respects, the individual parameter value storage unit 222 is similar to the individual parameter value storage unit 122.

[0109] The function calculation unit 221 specifies a function by setting the parameter values ​​acquired by the parameter value acquisition unit 142 from the common parameter value storage unit 210 and the parameter values ​​stored in the individual parameter value storage unit 222 as parameters of a function template. Then, the function calculation unit 221 inputs the value from the node calculation unit 111 to the specified function to calculate a function value. The function calculation unit 221 outputs the calculated function value to the node calculation unit 111.

[0110] Comparing the function calculation unit 221 with the function calculation unit 121 of the parameter value common model calculation device 14, the party from which the parameter values ​​are acquired is different. The function calculation unit 121 acquires parameter values ​​from the common parameter value storage unit 130 via the parameter value acquisition unit 142. In contrast, the function calculation unit 221 acquires parameter values ​​from the common parameter value storage unit 210 via the parameter value acquisition unit 142, and acquires the remaining parameter values ​​from the individual parameter value storage unit 222. In other respects, the function calculation unit 221 is similar to the function calculation unit 121. As in the case of the function calculation units 121 of the parameter value common model calculation device 14, all the function calculation units 221 may each have one function template with the same content, but the function templates provided may differ depending on the function calculation units 221. As long as the same parameter values ​​can be set to the parameters of the function templates of the function calculation units 221 included in the same group, the functions specified by the parameter value settings may differ depending on the function calculation units 221.

[0111] In this way, the parameter value mixed model calculation device 24 stores the parameter values ​​to be set for the parameters of the function template separately in the common parameter value storage unit 210 and the individual parameter value storage unit 222. This allows the parameter value mixed model calculation device 24 to achieve both a reduction in storage capacity and high-precision processing. Specifically, in the parameter value mixed model calculation device 24, the common parameter values ​​stored in the common parameter value storage unit 210 are shared by the multiple mixing method weighting units 220, so that a relatively small memory capacity is required. Moreover, in the parameter value mixed model calculation device 24, by providing an individual parameter value storage unit 222 for each mixing method weighting unit 220 and storing parameter values ​​therein, it is possible to specify a different function for each mixing method weighting unit 220, and in this respect, it is expected that processing can be performed with high accuracy.

[0112] Instead of the table common model calculation device 17, the information processing system 1 may be provided with a device in which only a part of the function table is common. 17 is a diagram showing an example of the configuration of the table mixed model calculation device 27. The table mixed model calculation device 27 corresponds to an example of a device in which only a part of the function table is made common.

[0113] 17, the table mixed model calculation device 27 includes a node calculation unit 111, a common table storage unit 230, and a mixed table method weight processing unit 240. The mixed table method weight processing unit 240 includes a table reference unit 241, a group ID storage unit 141, and an individual table storage unit 242. The table mixing model calculation device 27 corresponds to an example of the neural network system 110. The mixing table type weight processing unit 240 corresponds to an example of the weight processing unit 112.

[0114] The common table storage unit 230 stores a function table for a part of the domain of the function. Comparing the common table storage unit 230 with the common table storage unit 150 of the table common model calculation device 17, the common table storage unit 150 stores a function table for the entire domain of the function, whereas the common table storage unit 230 stores a function table for only a part of the domain of the function. In other respects, the common table storage unit 230 is similar to the common table storage unit 150. As in the case of the common table storage unit 150, the function table stored in the common table storage unit 230 is also referred to as a common table.

[0115] The group ID storage unit 141 of the table mixed model calculation device 27 stores group IDs, similar to the group ID storage unit 141 of the table common model calculation device 17. In the table mixed model calculation device 27, the group ID identifies a group of the mixed table scheme weight processing unit 240. The individual table storage unit 242 stores a function table for a part of the domain of the function. Specifically, the individual table storage unit 242 stores a function table for an interval of the domain of the function other than the interval for which the common table storage unit 230 stores a function table.

[0116] The table reference unit 241 selects either the common table storage unit 230 or the individual table storage unit 242 to refer to the function table. Specifically, the table reference unit 241 knows the interval in the domain of the function in which the common table storage unit 230 stores the function table and the interval in which the individual table storage unit 242 stores the function table. When a value is input from the node operation unit 111, the table reference unit 241 selects either the common table storage unit 230 or the individual table storage unit 242, which stores a function table for the value from the node operation unit 111. Then, the table reference unit 241 reads out a function value when the argument value is a value from the node operation unit 111 from the selected one of the common table storage unit 230 or the individual table storage unit 242. The table reference unit 241 outputs the obtained function value to the node operation unit 111.

[0117] When referring to the common table in the common table storage unit 230, the table reference unit 241 reads out the group ID from the group ID storage unit 141. Then, the table reference unit 241 refers to the common table associated with the group ID read out from the group ID storage unit 141, among the common tables stored in the common table storage unit 230.

[0118] Comparing the table reference unit 241 with the table reference unit 161 of the table common model calculation device 17, the function table reference destination is different. The function calculation unit 121 refers to the common table stored in the common table storage unit 150 for the entire domain of the function. In contrast, the table reference unit 241 selects either the common table storage unit 230 or the individual table storage unit 242 according to the argument value of the function to refer to the function table. In other respects, the table reference unit 241 is similar to the table reference unit 161.

[0119] In this way, the table mixed model calculation device 27 stores the function tables separately in the common table storage unit 230 and the individual table storage unit 242. This allows the table mixed model calculation device 27 to achieve both a reduction in storage capacity and high-precision processing. Specifically, in the table mixed model calculation device 27, a common table stored in the common table storage unit 230 is shared by a plurality of mixed table-based weight processing units 240, so that a relatively small storage capacity is required. Also, in the table mixed model calculation device 27, by providing an individual table storage unit 242 for each mixed table-based weight processing unit 240 and storing a part of the function table, it is possible to specify a different partial function for each mixed table-based weight processing unit 240, and in this respect, it is expected that processing can be performed with high accuracy.

[0120] For example, by the parameter value mixed model calculation device 24 storing the parameter values ​​separately in the common parameter value memory unit 210 and the individual parameter value memory unit 222, depending on the function template, it is possible that the domain of the function is divided into an interval that is a partial function common to all mixed method weighting processors 220 in the same group, and an interval where a partial function is set for each mixed method weighting processor 220. In this case, for an interval of the domain of the function that is a partial function common to all the mixing method weighting processors 220 in the same group, the common table storage unit 230 of the table mixing model calculation device 27 may store a function table. Also, for an interval in which a partial function is set for each mixing method weighting processor 220, the individual table storage unit 242 of the table mixing model calculation device 27 may store a function table.

[0121] The information processing system 1 may include, instead of the table common model calculation device 17, a device that uses both a function template and a function table. 18 is a diagram showing an example of the configuration of the parameter table combined model calculation device 28. The parameter table combined model calculation device 28 corresponds to an example of a device that uses both a function template and a function table.

[0122] 18, the parameter table combination model calculation device 28 includes a node calculation unit 111, a common table storage unit 150, and a combination method weight processing unit 250. The combination method weight processing unit 250 includes a function calculation unit 251, a group ID storage unit 141, and an individual parameter value storage unit 122. The parameter table combination model calculation device 28 corresponds to an example of the neural network system 110. The combination method weighting processor 250 corresponds to an example of the weighting processor 112.

[0123] The common table storage unit 150 stores the common table in association with a group ID, as in the case of the table common model calculation device 17. In the case of the parameter table combined model calculation device 28, the common table storage unit 150 stores, for each group of the combined use method weight processing unit 250, a group ID indicating that group and a common table for that group. The common table storage unit 150 stores a common table for the entire domain of the function.

[0124] The group ID storage unit 141 stores group IDs, similar to the group ID storage unit 141 of the table common model calculation device 17. In the combination scheme weight processing unit 250, the group ID identifies the group of the combination scheme weight processing unit 250. The individual parameter value storage unit 122 stores the parameter values ​​of the function, similarly to the individual parameter value storage unit 122 of the parameter value individual model calculation device 11 .

[0125] Function calculation section 251 performs calculation of a function by using both a function indicated by the common table in common table storage section 150 and a function obtained by setting parameter values ​​to parameters of a function template. Specifically, the function calculation unit 251 reads out a function value from a common table, similar to the table reference unit 161 of the table common model calculation device 17. The function calculation unit 251 identifies a common table linked to a group ID stored in the group ID storage unit 141, among the common tables stored in the common table storage unit 150. Then, the function calculation unit 251 reads out a function value from the identified common table when the value from the node calculation unit 111 is the argument value. A function indicated by a common table associated with a group ID stored in group ID storage unit 141 is also referred to as a common function.

[0126] Furthermore, like the function calculation unit 121 of the parameter value individual model calculation device 11, the function calculation unit 251 specifies a function by setting the parameter values ​​stored in the individual parameter value storage unit 122 to the parameters of the function template that the function calculation unit 121 itself has. A function specified by setting the parameter values ​​stored in the individual parameter value storage unit 122 to the parameters of a function template is also called an individual function.

[0127] The function calculation unit 251 performs an adjustment on the common function using the individual function. For example, the function calculation unit 251 may perform the adjustment by using the individual function as a function indicating an offset value for each argument value. In this case, the function calculation unit 251 adds together a function value read from the common table and a function value calculated by inputting a value from the node calculation unit 111 to the individual function.

[0128] Alternatively, the function calculation unit 251 may perform the adjustment by using the individual function as a function indicating the coefficient value for each argument value. In this case, the function calculation unit 251 multiplies the function value read from the common table by the function value calculated by inputting the value from the node calculation unit 111 to the individual function.

[0129] Alternatively, the function calculation unit 251 may synthesize functions by using the individual functions as post-processing for the common function. In this case, the function calculation unit 251 inputs the function value read from the common table to the individual function to calculate the function value. Alternatively, the function calculation unit 251 may synthesize functions by using the individual functions as pre-processing for the common function. In this case, the function calculation unit 251 inputs the value from the node calculation unit 111 to the individual function to calculate a function value. Then, the function calculation unit 251 refers to the common table linked to the group ID stored in the group ID storage unit 141, and reads out the function value when the function value obtained by the individual function is input to the common function.

[0130] According to the parameter table combination model calculation device 28, a complex function can be treated as a common function, and the function can be adjusted for each combination method weighting processor 250 by using an individual function. By storing the common function in the form of a function table in the common table storage unit 150, the function calculation unit 251 can obtain the function value with a relatively light load and at a relatively high speed even if the common function is a complex function. In addition, by storing a common table for each group of the combination method weight processing unit 250 in the common table storage unit 150, the memory capacity for storing the common table can be relatively small.

[0131] As the individual function, it is possible to use a simple function such as a ramp function, a linear function, an exponential function, or a logarithmic function. By using a simple function as the individual function, the memory capacity required for storing the parameter values ​​in the individual parameter value storage unit 122 can be small. In addition, by using a simple function as the individual function, the function calculation unit 251 can calculate the function value with a relatively light load and at a relatively high speed.

[0132] The target of the function to be shared is not limited to the function for weighting signals at the edges of the neural network, and may be, for example, the activation function of the neural network. 19 is a diagram showing an example of the configuration of a neural network system 34. The neural network system 34 corresponds to an example of the neural network system 110. A common method node operation unit 260 corresponds to an example of the node operation unit 111.

[0133] 19, of the components of the neural network system 34, the common parameter value storage unit 130 and the common method node calculation unit 260 are shown. The common method node calculation unit 260 includes an adder 261 and an activation function calculation unit 262. The activation function calculation unit 262 includes a function calculation unit 121, a group ID storage unit 141, and a parameter value acquisition unit 142. The adder 261 sums up the input values ​​to the common method node calculator 260 . The activation function calculation unit 262 calculates a function value when the output of the addition unit 261 is input to the activation function.

[0134] The common parameter value storage unit 130, the group ID storage unit 141, the parameter value acquisition unit 142, and the function calculation unit 121 are similar to the common parameter value storage unit 130, the group ID storage unit 141, the parameter value acquisition unit 142, and the function calculation unit 121 of the parameter value common model calculation device 14, except that the function to be calculated is an activation function.

[0135] 19, the common method node operation units 260 are divided into groups, and an activation function is set for each group. Therefore, in the neural network system 34, a different activation function can be used for each group of the common method node operation units 260. Moreover, the activation function changes by changing the common parameter value stored in the common parameter value storage unit 130. The activation function may be a learning target. For example, in the same way that the parameter value commonization device sets the parameter value common model calculation device 14 based on the trained parameter value individual model calculation device 11, after the activation function is set by learning a neural network capable of setting an activation function for each node, grouping may be performed based on the similarity of the activation functions, and the neural network system 34 may be set based on the grouping result. Alternatively, functions of a model other than a neural network may be the subject of commonality.

[0136] As described above, the function calculation unit 121 performs calculations using functions. The common function information storage unit stores values ​​that define functions for each group into which the function calculation unit 121 is divided. According to the arithmetic device, since the values ​​that define the function are stored for each group of function arithmetic units, the memory capacity for storing the values ​​for defining the function can be made relatively small.

[0137] The common parameter value storage unit 130 corresponds to an example of a common function information storage unit. The parameter values ​​correspond to an example of values ​​that define a function. The parameter value common model calculation device 14 corresponds to an example of a calculation device. Moreover, the common table storage unit 150 corresponds to an example of a common function information storage unit. The argument values ​​and function values ​​of the function table correspond to an example of values ​​that define a function. The table common model calculation device 17 corresponds to an example of a calculation device.

[0138] Furthermore, the function calculation unit 121 performs calculations using functions indicated by parameter values ​​set in a function template having parameters. The common parameter value storage unit 130 stores, for each group of the function calculation units 121, parameter values ​​common to the function calculation units 121 included in that group. According to the parameter value common model calculation device 14, the parameter values ​​of the function are stored for each group of the function calculation section 121, so that the storage capacity for storing the parameter values ​​can be made relatively small.

[0139] Moreover, the common table storage unit 150 stores a function table in which argument values ​​of a function are associated with function values ​​for each group of the function calculation unit 121. The function calculation unit 121 performs calculations using functions by referring to the function table. According to the table common model calculation device 17, since the function table is stored for each group of the function calculation units 121, the storage capacity for storing the function table can be made relatively small.

[0140] Moreover, the individual function information storage unit stores some of the set of values ​​defining one function for each function calculation unit 121. The common function information storage unit stores values ​​of the set of values ​​defining the one function other than the values ​​stored in the individual function information storage unit. According to the calculation device, a function can be identified for each number calculation unit in that a portion of a set of values ​​defining a function is stored for each function calculation unit 121, and the memory capacity for storing values ​​for defining a function can be made relatively small in that a portion of a set of values ​​defining a function is stored for each group of function calculation units 121. The individual parameter value storage unit 222 corresponds to an example of an individual function value storage unit. The common parameter value storage unit 210 corresponds to an example of a common function information storage unit. The parameter value mixed model calculation device 24 corresponds to an example of a calculation device. Moreover, the individual table storage unit 242 corresponds to an example of an individual function value storage unit, the common table storage unit 230 corresponds to an example of a common function information storage unit, and the table mixed model calculation device 27 corresponds to an example of a calculation device.

[0141] Furthermore, the function calculation unit 121 weights the signals at the edges of the neural network. The arithmetic unit allows the use of relatively complex functions for weighting signals, which is expected to enable the number of nodes in the neural network to be relatively small.

[0142] Furthermore, the function calculation unit 121 calculates the activation function at the node of the neural network. The computing device allows activation functions to be set for each group of nodes, which is expected to allow the number of nodes in the neural network to be relatively small.

[0143] The grouping unit 311 groups the first function calculation units of the parameter value individual model calculation device 11, which includes a plurality of first function calculation units that perform calculations using functions indicated by parameter values ​​set in a function template having parameters, and an individual parameter value storage unit that stores parameter values ​​for each of the first function calculation units, based on the similarity of the functions for each of the first function calculation units. The common parameter value calculation unit 312 calculates a common parameter value, which is a parameter value of one function used instead of the functions in all of the first function calculation units included in each group of the first function calculation units, for each group of the first function calculation units. The parameter value common model setting unit 313 sets groups according to the groups of the first function calculation units for the second function calculation units of the parameter value common model calculation device 14, which includes a plurality of second function calculation units that perform calculations using functions having parameters, and a common parameter value storage unit 130 that stores parameter values ​​for each group into which the second function calculation units are grouped, and stores the common parameter value in the common parameter value storage unit 130.

[0144] According to the parameter value commonization device 13, it is possible to generate a parameter value common model calculation device 14 in which the functions for each first function calculation section of the parameter value individual model calculation device 11 are commonized for each group based on the similarity of the functions, thereby reducing the memory capacity required to store the parameter values ​​of the functions.

[0145] The table calculation unit 321 calculates a function table for each group in which argument values ​​of a function are linked to function values ​​by parameter values ​​stored in the common parameter value storage unit 130 of the parameter value common model calculation device 14, which includes a plurality of function calculation units 121 that perform calculations using a function indicated by setting parameter values ​​in a function template having parameters, and a common parameter value storage unit 130 that stores parameter values ​​for each group into which the function calculation units 121 are grouped. The table common model setting unit 322 sets a group for the table reference unit 161 according to the group of the function calculation unit 121 of the table common model calculation device 17, which includes a plurality of table reference units 161 that perform calculations by referring to a function table in which argument values ​​of a function are linked to function values, and a common table storage unit 150 that stores a function table for each group into which the table reference units 161 are grouped, and stores the function table calculated for each group by the table calculation unit 321 in the common table storage unit 150. According to the tabulation device 16, it is possible to generate a table common model calculation device 17 in which the function indicated by the common parameters of the parameter value common model calculation device 14 is replaced with a function table, and it is expected that the calculation speed will be increased by tabulating the functions.

[0146] <Second embodiment> Fig. 20 is a diagram showing an example of a device configuration of an information processing system according to the second embodiment. In the configuration shown in Fig. 20, an information processing system 2 includes a parameter value common model calculation device 14, a group setting method learning control device 25, a tabulation device 16, and a table common model calculation device 17. The parameter value common model calculation device 14, the tabulation device 16, and the table common model calculation device 17 are similar to those in the configuration of FIG. 1, and detailed description thereof will be omitted here.

[0147] Comparing the configuration of the information processing system 2 with the configuration of the information processing system 1, the information processing system 2 does not include the parameter value individual model calculation device 11, the parameter value individual learning control device 12, and the parameter value commonization device 13. Moreover, the information processing system 2 includes a group setting method learning control device 25 instead of the parameter value common learning control device 15.

[0148] In the information processing system 1, the parameter value commonizing device 13 sets the parameter value common model computing device 14 based on the setting of the learned parameter value individual model computing device 11. In contrast, in the information processing system 2, the parameter value common model computing device 14 performs learning under the control of the group setting method learning control device 25. The parameter value common model computing device 14 updates the common parameter values ​​through learning.

[0149] Fig. 21 is a diagram showing an example of the configuration of group setting method learning control device 25. In the configuration shown in Fig. 21, group setting method learning control device 25 includes a temporary group setting unit 331, a learning control unit 332, a grouping determination unit 333, and a parameter value common model setting unit 313. The group setting method learning control device 25 corresponds to an example of a common calculation setting device.

[0150] The group provisional setting section 331 divides the function calculation sections 121 of the parameter value common model calculation device 14 into a plurality of groups. The learning control unit 332 controls the parameter value common model calculation device 14 to learn the common parameter values ​​in the common parameter value storage unit for each grouping by the group temporary setting unit 331. That is, the learning control unit 332 causes the parameter value common model calculation device 14 to perform learning and update the common parameter values ​​through the learning.

[0151] The learning control unit 332 may store, for each learning for each grouping, the grouping of the common scheme weight processing unit 140 and the common parameter value for each group stored in the common parameter value storage unit 130 upon completion of learning. The learning control unit 332 may store, as the grouping of the common scheme weight processing unit 140, the group IDs stored in each group ID storage unit 141.

[0152] The grouping determination unit 333 selects one of the groupings based on the learning results for each grouping. For example, the grouping determination unit 333 may evaluate the learning results using an evaluation function that gives a higher evaluation as the processing accuracy of the parameter value common model calculation device 14 is higher and that gives a higher evaluation as the number of common parameter values ​​stored in the common parameter value storage unit 130 is smaller. Then, the grouping determination unit 333 may select the grouping with the highest evaluation of the learning results.

[0153] The parameter value common model setting unit 313 sets groups according to the selected grouping for the function calculation unit 121 of the parameter value common model calculation device 14. Specifically, the parameter value common model setting unit 313 stores, for each common method weight processing unit 140, in the group ID storage unit 141, the group ID of the group including the common method weight processing unit 140. Furthermore, the parameter value common model setting unit 313 stores the common parameter values ​​obtained as a result of learning in the selected grouping in the common parameter value storage unit.

[0154] FIG. 22 is a flowchart showing an example of a processing procedure when the information processing system 2 performs setting of the parameter value common model computing device 14 and the table common model computing device 17. In the processing of FIG. 22, the group setting method learning control device 25 controls the parameter value common model calculation device 14 to perform learning for each of the multiple groupings of the common method weight processing unit 140, and sets the parameter value common model calculation device 14 based on the learning results (step S211). The parameter value common model calculation device 14 may be operated by, for example, mounting the parameter value common model calculation device 14 on a product.

[0155] Next, the tabulation device 16 creates a function table of the function indicated by the common parameter values ​​stored in the common parameter value storage unit 130 of the parameter value common model calculation device 14, and sets the function tabulation result in the table common model calculation device 17 (step S212). The process in step S212 is similar to the process in step S114 in FIG. The table common model calculation device 17 may be operated by, for example, mounting the table common model calculation device 17 on a product. After step S212, the information processing system 2 ends the process of FIG.

[0156] 23 is a flowchart showing an example of a processing procedure in which the group setting method learning control device 25 controls the parameter value common model calculation device 14 to perform learning. The group setting method learning control device 25 performs the processing of FIG. 23 in step S211 of FIG. In the process of FIG. 23, the group provisional setting unit 331 divides the common method weight processing units 140 of the parameter value common model calculation device 14 into groups, and sets the groups of the common method weight processing units 140 (S221).

[0157] The grouping here is a provisional grouping for trying various groupings. The provisional group setting unit 331 may randomly set the groups of the common scheme weight processing unit 140. Alternatively, the provisional group setting unit 331 may perform all possible groupings for the common scheme weight processing unit 140. The temporary group setting unit 331 stores a group ID indicating the group of the common scheme weight processing unit 140 in the group ID storage unit 141, thereby setting the group of the common scheme weight processing unit 140.

[0158] Next, the learning control unit 332 performs initial settings for learning of the parameter value common model calculation device 14 (step S222). For example, the learning control unit 332 causes the common parameter value storage unit 130 to store initial values ​​of the common parameter values ​​for each group. Then, the learning control unit 332 controls the parameter value common model calculation device 14 to perform learning (step S223). The parameter value common model calculation device 14 updates the common parameter values ​​through learning.

[0159] When the learning of the parameter value common model calculation device 14 is completed, the grouping determination unit 333 evaluates the learning result (step S224). For example, the grouping determination unit 333 may calculate and store an evaluation value of the learning result for each grouping of the common method weight processing unit 140 using a predetermined evaluation function.

[0160] Next, the learning control unit 332 determines whether or not a condition for ending the repetition of learning by the parameter value common model calculation device 14 is met (step S225). The termination condition here is not limited to a specific condition. For example, the termination condition in step S225 may be whether grouping has been performed a predetermined number of times or more. Alternatively, the termination condition in step S225 may be whether all possible groupings have been performed by the common scheme weight processor 140.

[0161] If the learning control unit 332 determines that the end condition is not satisfied (step S225: NO), the process returns to step S221. On the other hand, if the learning control unit 332 determines that the end condition is met (step S225: YES), the grouping determination unit 333 selects one of the groupings based on the learning result (step S226). For example, the grouping determination unit 333 selects the grouping with the largest evaluation value (highest evaluation) calculated in step S224.

[0162] Next, the parameter value common model setting unit 313 sets the parameter value common model calculation device 14 according to the grouping selected by the grouping determination unit (step S227). Specifically, the parameter value common model setting unit 313 causes the group IDs stored at the time of the selected grouping to be stored again in each of the group ID storage units 141. In addition, the parameter value common model setting unit 313 causes the common parameter value storage unit 130 to store again the common parameter values ​​obtained by learning the selected grouping. After step S227, the group setting method learning control device 25 ends the process of FIG.

[0163] As in the case of the information processing system 1, the information processing system 2 may be provided with a parameter value mixed model calculation device 24 instead of the parameter value common model calculation device 14. The information processing system 2 may be provided with a table mixed model calculation device 27 instead of the table common model calculation device 17. As in the case of the information processing system 1, the target of the common function in the information processing system 2 is not limited to the function for weighting signals at the edges of the neural network. For example, as in the neural network system 34 (FIG. 19), the activation function of the neural network may be the target of the common function. Alternatively, functions of a model other than a neural network may be the subject of commonality.

[0164] As described above, the group tentative setting unit 331 performs a plurality of groupings of the function calculation units 121 of the parameter value common model calculation device 14, which includes a plurality of function calculation units 121 that perform calculations using functions having parameters, and a common parameter value storage unit 130 that stores parameter values ​​for each group into which the function calculation units 121 are grouped. The learning control unit 332 causes the parameter value common model calculation device to learn the parameter values ​​for each group of the function calculation units 121 for each grouping. The grouping determination unit 333 selects one of the groupings based on the learning results for each grouping. The parameter value common model setting unit 313 sets groups according to the selected grouping for the function calculation units 121 of the parameter value common model calculation device 14, and stores the parameter values ​​in the learning results for the selected grouping in the common parameter value storage unit 130. According to the group setting type learning control device 25, the parameter value common model calculation device 14 can be set even if the parameter value individual model calculation device 11 is not present.

[0165] 24 is a schematic block diagram showing the configuration of a computer according to at least one embodiment. In the configuration shown in FIG. 10, a computer 700 includes a CPU (Central Processing Unit) 710, a main memory device 720, an auxiliary memory device 730, and an interface 740.

[0166] Any one or more of the parameter value individual model calculation device 11, the parameter value individual learning control device 12, the parameter value commonization device 13, the parameter value common model calculation device 14, the parameter value common learning control device 15, the tabulation device 16, the table common model calculation device 17, the parameter value mixed model calculation device 24, the table mixed model calculation device 27, and the group setting method learning control device 25 may be implemented in the computer 700. In this case, the operation of each of the above-mentioned processing units is stored in the auxiliary storage device 730 in the form of a program. The CPU 710 reads the program from the auxiliary storage device 730, expands it in the main storage device 720, and executes the above-mentioned processing according to the program. The CPU 710 also secures a storage area corresponding to each of the above-mentioned storage units in the main storage device 720 according to the program. The auxiliary storage device 730 is, for example, a non-transitory recording medium such as a CDC (Compact Disc) or a DVD (digital versatile disc).

[0167] When the parameter value individual model calculation device 11 is implemented in a computer 700, the node calculation unit 111, the individual scheme weight processing unit 120, and the operations of each of these units are stored in the form of a program in an auxiliary storage device 730. The CPU 710 reads the program from the auxiliary storage device 730, loads it into the main storage device 720, and executes the above-mentioned processing according to the program. Furthermore, the CPU 710 reserves a storage area in the main storage device 720 corresponding to the individual parameter value storage unit 122 in accordance with the program. Communication between the parameter value individual model calculation device 11 and other devices is performed by the interface 740 having a communication function and performing communication under the control of the CPU 710. Interaction between the parameter value individual model calculation device 11 and a user is performed by the interface 740 having a display device and an input device, displaying various images under the control of the CPU 710, and accepting user operations.

[0168] When the parameter value individual learning control device 12 is implemented in the computer 700, the operation of the parameter value individual learning control device 12 is stored in the form of a program in the auxiliary storage device 730. The CPU 710 reads the program from the auxiliary storage device 730, loads it in the main storage device 720, and executes the above-mentioned processing in accordance with the program. Furthermore, the CPU 710 reserves a storage area in the main storage device 720 for the parameter value individual learning control device 12 to carry out processing in accordance with the program. Communication between the parameter value individual learning control device 12 and other devices is performed by the interface 740 having a communication function and performing communication under the control of the CPU 710. Interaction between the parameter value individual learning control device 12 and a user is performed by the interface 740 having a display device and an input device, displaying various images under the control of the CPU 710, and accepting user operations.

[0169] When the parameter value commonization device 13 is implemented in a computer 700, the operations of the grouping unit 311, the common parameter value calculation unit 312, and the parameter value common model setting unit 313 are stored in the form of programs in the auxiliary storage device 730. The CPU 710 reads the programs from the auxiliary storage device 730, loads them in the main storage device 720, and executes the above-mentioned processes in accordance with the programs. Furthermore, the CPU 710 reserves a storage area in the main storage device 720 for the parameter value common device 13 to perform processing in accordance with the program. Communication between the parameter value common device 13 and other devices is performed by the interface 740 having a communication function and performing communication under the control of the CPU 710. Interaction between the parameter value common device 13 and a user is performed by the interface 740 having a display device and an input device, displaying various images under the control of the CPU 710, and accepting user operations.

[0170] When the parameter value common model calculation device 14 is implemented in a computer 700, the node calculation unit 111, the common method weight processing unit 140, and the operations of each of these units are stored in the form of a program in the auxiliary storage device 730. The CPU 710 reads the program from the auxiliary storage device 730, loads it into the main storage device 720, and executes the above-mentioned processing according to the program. Furthermore, the CPU 710 allocates storage areas in the main storage device 720 corresponding to the common parameter value storage unit 130 and the group ID storage unit 141 in accordance with the program. Communication between the parameter value common model calculation device 14 and other devices is performed by the interface 740 having a communication function and performing communication under the control of the CPU 710. Interaction between the parameter value common model calculation device 14 and a user is performed by the interface 740 having a display device and an input device, displaying various images under the control of the CPU 710, and accepting user operations.

[0171] When the parameter value common learning control device 15 is implemented in the computer 700, the operation of the parameter value common learning control device 15 is stored in the form of a program in the auxiliary storage device 730. The CPU 710 reads the program from the auxiliary storage device 730, loads it in the main storage device 720, and executes the above-mentioned processing in accordance with the program. Furthermore, the CPU 710 reserves a storage area in the main storage device 720 for the parameter value common learning control device 15 to perform processing in accordance with the program. Communication between the parameter value common learning control device 15 and other devices is performed by the interface 740 having a communication function and performing communication under the control of the CPU 710. Interaction between the parameter value common learning control device 15 and a user is performed by the interface 740 having a display device and an input device, displaying various images under the control of the CPU 710, and accepting user operations.

[0172] When the tabulation device 16 is implemented in a computer 700, the operations of the table calculation unit 321 and the table common model setting unit 322 are stored in the form of a program in the auxiliary storage device 730. The CPU 710 reads the program from the auxiliary storage device 730, loads it in the main storage device 720, and executes the above-mentioned processing according to the program. Furthermore, the CPU 710 reserves a storage area in the main storage device 720 for the tabulation device 16 to carry out processing in accordance with the program. Communication between the tabular device 16 and other devices is performed by the interface 740 having a communication function and performing communication under the control of the CPU 710. Interaction between the tabular device 16 and a user is performed by the interface 740 having a display device and an input device, displaying various images under the control of the CPU 710, and accepting user operations.

[0173] When the table common model calculation device 17 is implemented in a computer 700, the node calculation unit 111, the common table method weight processing unit 160, and the operations of each of these units are stored in the form of a program in an auxiliary storage device 730. The CPU 710 reads the program from the auxiliary storage device 730, loads it in the main storage device 720, and executes the above-mentioned processing according to the program. Furthermore, the CPU 710 allocates storage areas in the main storage device 720 corresponding to the common table storage unit 150 and the group ID storage unit 141 in accordance with the program. Communication between the table common model calculation device 17 and other devices is performed by the interface 740 having a communication function and performing communication under the control of the CPU 710. Interaction between the table common model calculation device 17 and a user is performed by the interface 740 having a display device and an input device, displaying various images under the control of the CPU 710 and accepting user operations.

[0174] When the parameter value mixed model calculation device 24 is implemented in a computer 700, the node calculation unit 111, the mixed method weighting processing unit 220, and the operations of each of these units are stored in the form of a program in an auxiliary storage device 730. The CPU 710 reads the program from the auxiliary storage device 730, loads it into the main storage device 720, and executes the above-mentioned processing according to the program. Furthermore, the CPU 710 allocates storage areas in the main storage device 720 corresponding to the common parameter value storage unit 210, the group ID storage unit 141, and the individual parameter value storage unit 222 in accordance with the program. Communication between the table common model calculation device 17 and other devices is performed by the interface 740 having a communication function and performing communication under the control of the CPU 710. Interaction between the table common model calculation device 17 and a user is performed by the interface 740 having a display device and an input device, displaying various images under the control of the CPU 710 and accepting user operations.

[0175] When the table mixed model calculation device 27 is implemented in a computer 700, the node calculation unit 111, the mixed table method weight processing unit 240, and the operations of each of these units are stored in the form of a program in the auxiliary storage device 730. The CPU 710 reads the program from the auxiliary storage device 730, loads it into the main storage device 720, and executes the above-mentioned processing according to the program. Furthermore, the CPU 710 reserves storage areas in the main storage device 720 corresponding to the common table storage unit 230, the group ID storage unit 141, and the individual table storage unit 242 in accordance with the program. Communication between the table common model calculation device 17 and other devices is performed by the interface 740 having a communication function and performing communication under the control of the CPU 710. Interaction between the table common model calculation device 17 and a user is performed by the interface 740 having a display device and an input device, displaying various images under the control of the CPU 710, and accepting user operations.

[0176] When the group setting method learning control device 25 is implemented in a computer 700, the operations of the temporary group setting unit 331, the learning control unit 332, the grouping determination unit 333, and the parameter value common model setting unit 313 are stored in the form of programs in the auxiliary storage device 730. The CPU 710 reads the programs from the auxiliary storage device 730, loads them into the main storage device 720, and executes the above-mentioned processes according to the programs. Furthermore, the CPU 710 reserves a storage area in the main storage device 720 for the group setting method learning control device 25 to perform processing in accordance with the program. Communication between the group setting method learning control device 25 and other devices is performed by the interface 740 having a communication function and performing communication under the control of the CPU 710. Interaction between the group setting method learning control device 25 and a user is performed by the interface 740 having a display device and an input device, displaying various images under the control of the CPU 710 and accepting user operations.

[0177] Note that the processing of each part may be performed by recording a program for implementing all or part of the functions of the parameter value individual model calculation device 11, the parameter value individual learning control device 12, the parameter value common device 13, the parameter value common model calculation device 14, the parameter value common learning control device 15, the tabulation device 16, the table common model calculation device 17, the parameter value mixed model calculation device 24, the table mixed model calculation device 27, and the group setting method learning control device 25 on a computer-readable recording medium, and reading and executing the program recorded on the recording medium into a computer system. Note that the term "computer system" here includes hardware such as an OS (Operating System) and peripheral devices. Furthermore, the term "computer-readable recording medium" refers to portable media such as flexible disks, optical magnetic disks, ROMs (Read Only Memory), and CD-ROMs (Compact Disc Read Only Memory), as well as storage devices such as hard disks built into computer systems. The above-mentioned program may be for realizing part of the above-mentioned functions, or may be capable of realizing the above-mentioned functions in combination with a program already recorded in the computer system.

[0178] Although an embodiment of the present invention has been described in detail above with reference to the drawings, the specific configuration is not limited to this embodiment, and design changes and the like that do not deviate from the gist of the present invention are also included.

[0179] Some or all of the above embodiments may be described as follows, but are not limited to the following:

[0180] (Appendix 1) A plurality of function calculation units that perform calculations using functions; a common function information storage unit that stores values ​​that define the functions for each group into which the function calculation units are grouped; A computing device comprising:

[0181] (Appendix 2) the function calculation unit performs a calculation using the function indicated by a function template having parameters and parameter values ​​set therein; the common function information storage unit stores, for each group, a parameter value common to the function calculation units included in the group; 2. The computing device of claim 1.

[0182] (Appendix 3) the common function information storage unit stores, for each group, a function table in which argument values ​​of the function are associated with function values; the function calculation unit performs a calculation using the function by referring to the function table; 2. The computing device of claim 1.

[0183] (Appendix 4) Further comprising an individual function information storage unit that stores a part of a set of values ​​that defines one function for each of the function calculation units; the common function information storage unit stores a value other than the value stored in the individual function information storage unit among a set of values ​​defining the one function; 4. A computing device according to any one of claims 1 to 3.

[0184] (Appendix 5) The function calculation unit weights signals at edges of the neural network. 5. A computing device according to any one of claims 1 to 4.

[0185] (Appendix 6) The function calculation unit calculates an activation function at a node of a neural network. 6. A computing device according to any one of claims 1 to 5.

[0186] (Appendix 7) a grouping unit that groups the first function calculation units of a parameter value individual model calculation device, the first function calculation units being provided with a plurality of first function calculation units that perform calculations using functions indicated by parameter values ​​set in a function template having parameters, and an individual parameter value storage unit that stores parameter values ​​for each of the first function calculation units, based on a similarity of the functions for each of the first function calculation units; a common parameter value calculation unit that calculates, for each group of the first function calculation units, a common parameter value that is a parameter value of one function that is used instead of the functions in all of the first function calculation units included in that group; a parameter value common model calculation device including a plurality of second function calculation units that perform calculations using functions having parameters, and a common parameter value storage unit that stores parameter values ​​for each group into which the second function calculation units are grouped; and a parameter value common model setting unit that sets groups for the second function calculation units according to the groups of the first function calculation units and stores the common parameter values ​​in the common parameter value storage unit; A common calculation setting device comprising:

[0187] (Appendix 8) a parameter value common model calculation device including a plurality of function calculation units that perform calculations using functions indicated by parameter values ​​set in a function template having parameters, and a common parameter value storage unit that stores parameter values ​​for each group into which the function calculation units are grouped; and a table calculation unit that calculates, for each group, a function table in which argument values ​​of the functions and function values ​​are associated with each other based on the parameter values ​​stored in the common parameter value storage unit; a table common model calculation device including a plurality of table reference units that perform calculations by referring to a function table in which argument values ​​of a function and function values ​​are associated with each other, and a common table storage unit that stores a function table for each group into which the table reference units are grouped; a table common model setting unit that sets groups corresponding to the groups of the function calculation units for the table reference units, and stores the function table calculated for each group by the table calculation unit in the common table storage unit; A common calculation setting device comprising:

[0188] (Appendix 9) a group provisional setting unit that groups the function calculation units in a plurality of ways in a parameter value common model calculation device that includes a plurality of function calculation units that perform calculations using functions having parameters, and a common parameter value storage unit that stores parameter values ​​for each group into which the function calculation units are grouped; a learning control unit that causes the parameter value common model calculation device to learn the parameter values ​​for each group of the function calculation unit for each of the groupings; A grouping determination unit that selects one of the groupings based on the learning results for each grouping; a parameter value common model setting unit that sets groups according to a selected grouping for the function calculation unit of the parameter value common model calculation device and stores parameter values ​​in the learning results for the selected grouping in the common parameter value storage unit; A common calculation setting device comprising:

[0189] (Appendix 10) The grouped function calculation units read out parameter values ​​from a common parameter value storage unit that stores parameter values ​​for each group; the function calculation unit performs calculation using a function indicated by setting parameter values ​​read from a common parameter value storage unit to a function template having parameters; A calculation method including:

[0190] (Appendix 11) a parameter value individual model calculation device including a plurality of first function calculation units that perform calculations using functions indicated by parameter values ​​set in a function template having parameters, and an individual parameter value storage unit that stores parameter values ​​for each of the first function calculation units; grouping the first function calculation units based on similarity of the functions for each of the first function calculation units; calculating, for each group of the first function calculation units, a common parameter value that is a parameter value of one function used in place of the functions in all of the first function calculation units included in that group; a parameter value common model calculation device including a plurality of second function calculation units that perform calculations using functions having parameters, and a common parameter value storage unit that stores parameter values ​​for each group into which the second function calculation units are grouped, the device setting groups for the second function calculation units according to the groups of the first function calculation units, and storing the common parameter values ​​in the common parameter value storage unit; A common calculation setting method including:

[0191] (Appendix 12) a parameter value common model calculation device including a plurality of function calculation units that perform calculations using functions indicated by parameter values ​​set in a function template having parameters, and a common parameter value storage unit that stores parameter values ​​for each group into which the function calculation units are grouped, calculating, for each group, a function table in which argument values ​​of the functions and function values ​​are associated with each other based on parameter values ​​stored in the common parameter value storage unit; a table common model calculation device including a plurality of table reference units that perform calculations by referring to a function table in which argument values ​​of a function and function values ​​are associated with each other, and a common table storage unit that stores a function table for each group into which the table reference units are grouped, wherein the table reference units are set to groups corresponding to the groups of the function calculation units, and the function table calculated for each group is stored in the common table storage unit; A common calculation setting method including:

[0192] (Appendix 13) a parameter value common model calculation device including a plurality of function calculation units that perform calculations using functions having parameters, and a common parameter value storage unit that stores parameter values ​​for each group into which the function calculation units are grouped, said function calculation units being grouped in a plurality of ways; calculating the parameter values ​​for each group of the function calculation unit by learning for each of the groupings; selecting one of the groupings based on the learning results for each of the groupings; setting a group according to the selected grouping for the function calculation unit of the parameter value common model calculation device, and storing parameter values ​​in the learning results for the selected grouping in the common parameter value storage unit; A common calculation setting method including:

[0193] (Appendix 14) On the computer, The grouped function calculation units read out parameter values ​​from a common parameter value storage unit that stores parameter values ​​for each group; the function calculation unit performs calculation using a function indicated by setting parameter values ​​read from a common parameter value storage unit to a function template having parameters; The program to be executed.

[0194] (Appendix 15) On the computer, a parameter value individual model calculation device including a plurality of first function calculation units that perform calculations using functions indicated by parameter values ​​set in a function template having parameters, and an individual parameter value storage unit that stores parameter values ​​for each of the first function calculation units; grouping the first function calculation units based on similarity of the functions for each of the first function calculation units; calculating, for each group of the first function calculation units, a common parameter value that is a parameter value of one function used in place of the functions in all of the first function calculation units included in that group; a parameter value common model calculation device including a plurality of second function calculation units that perform calculations using functions having parameters, and a common parameter value storage unit that stores parameter values ​​for each group into which the second function calculation units are grouped, the device setting groups for the second function calculation units according to the groups of the first function calculation units, and storing the common parameter values ​​in the common parameter value storage unit; A program for executing the above.

[0195] (Appendix 16) On the computer, a parameter value common model calculation device including a plurality of function calculation units that perform calculations using functions indicated by parameter values ​​set in a function template having parameters, and a common parameter value storage unit that stores parameter values ​​for each group into which the function calculation units are grouped, calculating, for each group, a function table in which argument values ​​of the functions and function values ​​are associated with each other based on parameter values ​​stored in the common parameter value storage unit; a table common model calculation device including a plurality of table reference units that perform calculations by referring to a function table in which argument values ​​of a function and function values ​​are associated with each other, and a common table storage unit that stores a function table for each group into which the table reference units are grouped, wherein the table reference units are set to groups corresponding to the groups of the function calculation units, and the function table calculated for each group is stored in the common table storage unit; A program for executing the above.

[0196] (Appendix 17) On the computer, a parameter value common model calculation device including a plurality of function calculation units that perform calculations using functions having parameters, and a common parameter value storage unit that stores parameter values ​​for each group into which the function calculation units are grouped, said function calculation units being grouped in a plurality of ways; calculating the parameter values ​​for each group of the function calculation unit by learning for each of the groupings; selecting one of the groupings based on the learning results for each of the groupings; setting a group according to the selected grouping for the function calculation unit of the parameter value common model calculation device, and storing parameter values ​​in the learning results for the selected grouping in the common parameter value storage unit; A program for executing the above. [Explanation of symbols]

[0197] 1, 2 Information processing system 11 Parameter value individual model calculation device 12 Parameter value individual learning control device 13 Parameter value commonization device 14 Parameter value common model calculation device 15 Parameter value common learning control device 16 Table making device 17 Table common model calculation device 24 Parameter value mixture model calculation device 25 Group setting type learning control device 27 Table Mixed Model Calculator 34, 110 Neural network system 111 Node Calculation Unit 112 Weighting Processor 120 Individual method weight processing unit 121, 221, 251 Function calculation section 122, 222 Individual parameter value storage unit 130, 210 Common parameter value storage section 140 Common method weight processing unit 141 Group ID storage section 142 Parameter value acquisition unit 150, 230 Common table storage section 160 Common table method weight processing unit 161, 241 Table Reference Section 220 Mixed method weight processing unit 240 Mixed table type weight processing unit 242 Individual table storage unit 250 Combined method weight processing unit 260 Common Method Node Calculation Unit 261 Addition section 262 Activation Function Calculation Unit 311 Grouping Department 312 Common parameter value calculation unit 313 Parameter Value Common Model Setting Unit 321 Table Calculation Unit 322 Table Common Model Setting Section 331 Group Provisional Setting Department 332 Learning control section 333 Grouping Decision Unit

Claims

1. A plurality of function calculation units that perform calculations using functions; a common function information storage unit that stores a part of a set of values ​​that defines the function for each group into which the function calculation units are grouped; an individual function information storage unit that stores, for each of the function calculation units, values ​​other than the values ​​stored in the common function information storage unit among a set of values ​​that define the function; A computing device comprising:

2. the function calculation unit performs a calculation using the function indicated by a function template having parameters and parameter values ​​set therein; the common function information storage unit stores, for each group, a parameter value common to the function calculation units included in the group as a value defining the function; The computing device of claim 1 .

3. the common function information storage unit stores, for each group, argument values ​​of the function, which are values ​​that define the function, and function values ​​in a function table in which the argument values ​​and the function values ​​are associated with each other; the function calculation unit performs a calculation using the function by referring to the function table; The computing device of claim 1 .

4. a grouping unit that groups the first function calculation units of a parameter value individual model calculation device, the first function calculation units being provided with a plurality of first function calculation units that perform calculations using functions indicated by parameter values ​​set in a function template having parameters, and an individual parameter value storage unit that stores parameter values ​​for each of the first function calculation units, based on a similarity of the functions for each of the first function calculation units; a common parameter value calculation unit that calculates, for each group of the first function calculation units, a common parameter value that is a parameter value of one function that is used in place of the functions in all of the first function calculation units included in that group; a parameter value common model calculation device including a plurality of second function calculation units that perform calculations using functions having parameters, and a common parameter value storage unit that stores parameter values ​​for each group into which the second function calculation units are grouped; a parameter value common model setting unit that sets groups for the second function calculation units according to the groups of the first function calculation units and stores the common parameter values ​​in the common parameter value storage unit; A common calculation setting device comprising:

5. a group provisional setting unit that groups the function calculation units in a plurality of ways in a parameter value common model calculation device that includes a plurality of function calculation units that perform calculations using functions having parameters, and a common parameter value storage unit that stores parameter values ​​for each group into which the function calculation units are grouped; a learning control unit that causes the parameter value common model calculation device to learn the parameter values ​​for each group of the function calculation unit for each of the groupings; A grouping determination unit that selects one of the groupings based on the learning results for each grouping; a parameter value common model setting unit that sets groups according to a selected grouping for the function calculation unit of the parameter value common model calculation device and stores parameter values ​​in the learning results for the selected grouping in the common parameter value storage unit; A common calculation setting device comprising:

6. a parameter value individual model calculation device including a plurality of first function calculation units that perform calculations using functions indicated by parameter values ​​set in a function template having parameters, and an individual parameter value storage unit that stores parameter values ​​for each of the first function calculation units; grouping the first function calculation units based on similarity of the functions for each of the first function calculation units; calculating, for each group of the first function calculation units, a common parameter value that is a parameter value of one function used in place of the functions in all of the first function calculation units included in that group; a parameter value common model calculation device including a plurality of second function calculation units that perform calculations using functions having parameters, and a common parameter value storage unit that stores parameter values ​​for each group into which the second function calculation units are grouped, the device setting groups for the second function calculation units according to the groups of the first function calculation units, and storing the common parameter values ​​in the common parameter value storage unit; A common calculation setting method including:

7. a parameter value common model calculation device including a plurality of function calculation units that perform calculations using functions having parameters, and a common parameter value storage unit that stores parameter values ​​for each group into which the function calculation units are grouped, said function calculation units being grouped in a plurality of ways; calculating the parameter values ​​for each group of the function calculation unit by learning for each of the groupings; selecting one of the groupings based on the learning results for each of the groupings; setting a group according to the selected grouping for the function calculation unit of the parameter value common model calculation device, and storing parameter values ​​in the learning results for the selected grouping in the common parameter value storage unit; A common calculation setting method including:

8. On the computer, a parameter value individual model calculation device including a plurality of first function calculation units that perform calculations using functions indicated by parameter values ​​set in a function template having parameters, and an individual parameter value storage unit that stores parameter values ​​for each of the first function calculation units; grouping the first function calculation units based on similarity of the functions for each of the first function calculation units; calculating, for each group of the first function calculation units, a common parameter value that is a parameter value of one function used in place of the functions in all of the first function calculation units included in that group; a parameter value common model calculation device including a plurality of second function calculation units that perform calculations using functions having parameters, and a common parameter value storage unit that stores parameter values ​​for each group into which the second function calculation units are grouped, the device setting groups for the second function calculation units according to the groups of the first function calculation units, and storing the common parameter values ​​in the common parameter value storage unit; A program for executing.

9. On the computer, a parameter value common model calculation device including a plurality of function calculation units that perform calculations using functions having parameters, and a common parameter value storage unit that stores parameter values ​​for each group into which the function calculation units are grouped, said function calculation units being grouped in a plurality of ways; calculating the parameter values ​​for each group of the function calculation unit by learning for each of the groupings; selecting one of the groupings based on the learning results for each of the groupings; setting a group according to the selected grouping for the function calculation unit of the parameter value common model calculation device, and storing parameter values ​​in the learning results for the selected grouping in the common parameter value storage unit; A program for executing.

Images