JP7421046B2 - Information acquisition device, information acquisition method and program - Google Patents

Description

The present invention relates to an information acquisition device, an information acquisition method, and a program.

Related to the interpretability of machine learning models, when image data is input to a neural network, which is a trained model, it is traced from the output node to the input node (in the opposite direction of the signal flow) and then fired. It is stated that the nodes that are currently being displayed are displayed.

Gregoire Montavona and 4 others, "Explaining nonlinear classification decisions with deep Taylor decomposition", Pattern Recognition 65, p.211-222, ELSEVIER, 2017

According to the method described in Non-Patent Document 1, it is possible to visualize the processing performed by a trained model on specific input data.
On the other hand, from the perspective of dealing with new input data, it is desirable to be able to interpret the model itself without relying on specific input data.

An object of the present invention is to provide an information acquisition device, an information acquisition method, and a program that can solve the above-mentioned problems.

According to a first aspect of the present invention, the information acquisition device includes an acquisition unit that calculates a distribution of node values with respect to a plurality of input data to a graphical model; Based on a value indicating the strength of the correlation of the distribution between each node on the path from the input node to the final node of the hidden layer, and a value indicating the degree of dispersion of the values transmitted from the final node to each output node. and a calculation unit that calculates an evaluation value of the degree of contribution of the route to the value of the output node .

According to a second aspect of the present invention, the information acquisition method includes the steps of determining the distribution of node values for a plurality of input data to the graphical model, calculating the correlation of the distribution between the nodes , and calculating the correlation of the distribution between the nodes. Based on a value indicating the strength of the correlation of the distribution between each node on the path from the input node to the final node of the hidden layer, and a value indicating the degree of dispersion of the values transmitted from the final node to each output node. , calculating an evaluation value of the degree of contribution of the route to the value of the output node .

According to the third aspect of the present invention, the program includes the steps of causing a computer to calculate the distribution of node values for a plurality of input data to the graphical model, calculating the correlation of the distribution between the nodes, and calculating the correlation of the distribution between the nodes , Based on a value indicating the strength of the correlation of the distribution between each node on the path from the input node to the final node of the hidden layer, and a value indicating the degree of dispersion of the values transmitted from the final node to each output node. and calculating an evaluation value of the degree of contribution of the route to the value of the output node .

According to this invention, information for interpreting a model can be provided.

FIG. 1 is a schematic block diagram showing a functional configuration of an information acquisition device according to an embodiment. It is a figure showing the example of a display of a target model by a display part concerning an embodiment. FIG. 6 is a diagram illustrating an example of display of the distribution of values of each node of the target model by the display unit according to the embodiment. FIG. 6 is a diagram illustrating a display example of a ranking of the degree of contribution of the path from the input node of the target model to the final node of the hidden layer to the degree of variation in the value of the output node, by the display unit according to the embodiment. 1 is a flowchart illustrating an example of a processing procedure when the information acquisition device according to the embodiment displays a ranking of the degree of contribution of the path from the input node of the target model to the final node of the hidden layer to the degree of variation in the value of the output node; be. FIG. 1 is a diagram illustrating an example of a configuration of an information acquisition device according to an embodiment. It is a figure showing an example of a processing procedure in an information acquisition method concerning an embodiment. FIG. 1 is a schematic block diagram showing the configuration of a computer according to at least one embodiment.

Hereinafter, embodiments of the present invention will be described, but the following embodiments do not limit the invention according to the claims. Furthermore, not all combinations of features described in the embodiments are essential to the solution of the invention.
FIG. 1 is a schematic block diagram showing the functional configuration of an information acquisition device according to an embodiment. With the configuration shown in the figure, the information acquisition device 100 includes a communication section 110, a display section 120, an operation input section 130, a storage section 180, and a control section 190. The storage unit 180 includes a model storage unit 181. The control unit 190 includes an acquisition unit 191 and a calculation unit 192.

The information acquisition device 100 provides information for interpreting a graphical model having a directed acyclic graph structure. Hereinafter, the model to be processed by the information acquisition device 100 (the model to which the information acquisition device 100 provides information) will also be referred to as a target model. The target model may be a feedforward neural network such as deep learning, but is not limited to this. Further, the target model is typically a trained model (a model obtained by machine learning), but is not limited to this.

The information acquisition device 100 is configured using a computer such as a personal computer (PC). Alternatively, the information acquisition device 100 may be configured using hardware dedicated to the information acquisition device 100, such as using an ASIC (Application Specific Integrated Circuit).

The communication unit 110 communicates with other devices. For example, the communication unit 110 receives input data for the target model from a computer such as a server device. However, the method by which the information acquisition device 100 acquires input data to the target model is not limited to a specific method.
The display unit 120 includes a display screen such as a liquid crystal panel or an LED (Light Emitting Diode) panel, and displays various images. For example, the display unit 120 displays information for interpreting the target model, which is calculated by the information acquisition device 100. However, the method by which the information acquisition device 100 outputs information is not limited to displaying on the display unit 120. For example, the communication unit 110 may transmit information to another device.

The operation input unit 130 includes input devices such as a keyboard and a mouse, and receives user operations. For example, when the operation input unit 130 receives a user operation instructing calculation of information for interpreting the target model, the information acquisition device 100 starts calculating information for interpreting the target model.

The storage unit 180 stores various data. The storage unit 180 is configured using a storage device included in the information acquisition apparatus 100.
The model storage unit 181 stores a target model. Note that FIG. 1 shows an example where the target model is executed by software. In this case, the control unit 190 reads the target model from the model storage unit 181 and performs processing related to the target model, such as computation on input data. However, the target model may be executed in hardware.

The control unit 190 controls each unit of the information acquisition device 100 to perform various processes. The functions of the control unit 190 are executed by, for example, a CPU (Central Processing Unit) included in the information acquisition device 100 reading and executing a program from the storage unit 180.
The acquisition unit 191 executes calculation of the target model for each of the plurality of input data. Thereby, the acquisition unit 191 calculates the value of each node of the target model for each of the plurality of input data. The acquisition unit 191 then obtains the distribution of node values for a plurality of input data for each node of the target model.

The calculation unit 192 calculates the correlation. Specifically, the calculation unit 192 calculates a correlation coefficient as a correlation between the distribution of node values between adjacent nodes. The adjacent nodes here are nodes at both ends of one edge. That is, adjacent nodes are two nodes that are directly connected by one edge without going through another node.
For example, the calculation unit 192 may calculate the correlation coefficient using a copula.

The calculation unit 192 further calculates an evaluation value of the degree of contribution (magnitude of influence) of the path from the input node of the target model to the final node of the hidden layer to the degree of variation in the value of the output node. The final node of the hidden layer here is a node adjacent to the output node among the nodes of the hidden layer.
Specifically, the calculation unit 192 calculates the product of the correlations of the distribution of node values between each node on the path from the input node of the target model to the final node of the hidden layer.

Further, the calculation unit 192 calculates a value indicating the degree of variation in the values transmitted from the final node to each output node. In the following, a case where the calculation unit 192 calculates the variance as a value indicating the degree of dispersion will be described as an example. However, the value indicating the degree of variation calculated by the calculation unit 192 is not limited to the variance. For example, the calculation unit 192 may calculate the standard deviation as a value indicating the degree of variation.
Then, the calculation unit 192 calculates the product of the correlation of the distribution of node values between each node on the path from the input node of the target model to the final node of the hidden layer, and the variance of the values transmitted from the final node to each output node. Calculate the product of

The calculation unit 192 calculates, for each of the plurality of paths from the input node of the target model to the final node of the hidden layer, an evaluation value of the degree of contribution of that path to the degree of variation in the value of the output node. For example, the calculation unit 192 calculates the evaluation value of the degree of contribution of each path to the degree of variation in the value of the output node for all paths from all input nodes to the final nodes of all hidden layers.
The calculation unit 192 then ranks the routes based on the evaluation value for each route. For example, the calculation unit 192 ranks routes in descending order of evaluation value. Alternatively, the calculation unit 192 may rank the routes in descending order of evaluation value.

FIG. 2 is a diagram showing an example of displaying a target model on the display unit 120. In the example of FIG. 2, the target model has an input layer, two hidden layers (middle layers), and an output layer. The input layer has four input nodes. The first hidden layer has six intermediate nodes. The second hidden layer has four intermediate nodes. The output layer has three output nodes.
However, the target model is not limited to the model with the structure shown in FIG.

From each node of the input layer to each node of the first layer of the hidden layer, from each node of the first layer of the hidden layer to each node of the second layer of the hidden layer, and from each node of the second layer of the hidden layer A directed edge is provided to each node of the output layer. A weighting coefficient is set for each edge, and the value output from a node is multiplied by the weighting coefficient at the edge and input to the next node.
In this way, the target model shown in FIG. 2 has a directed and acyclic graph structure.
Further, in FIG. 2, among the routes from the input node x_0 to the output node pred_2, a route via intermediate nodes h0_4 and h1_0 is shown.

FIG. 3 is a diagram showing a display example of the distribution of values of each node of the target model on the display unit 120. FIG. 3 shows an example of the distribution of values of each node of the target model shown in FIG. 2.
In the example of FIG. 3, the display unit 120 shows a histogram (frequency distribution diagram) for each node. The horizontal axis of each histogram indicates the values that the node can take with respect to a plurality of input data. The vertical axis indicates a value obtained by normalizing the number of times the node takes that value divided by the number of input data. This normalization allows the frequency distribution to be treated as a probability distribution.

Further, in the example of FIG. 3, the display unit 120 displays the frequency distribution of each node in a different pattern for each class classification result. This class classification is indicated by the value of the output node. In the example of FIG. 3, the nodes are classified into three classes: a pred_0 class, a pred_1 class, and a pred_2 class, depending on which one of the three nodes pred_0, pred_1, and pred_2 shows the largest value.

In order to display the class classification of the frequency distribution as shown in FIG. Label. The acquisition unit 191 calculates the value of each node for all the input data given, and then aggregates the values taken by the node for each node to create a histogram. At this time, the acquisition unit 191 counts the number of times a certain node takes a certain value for each class labeled with the value of the node. Thereby, the display unit 120 can display the frequency distribution of each node in a different pattern for each class classification result, as in the example of FIG.

As in the example of FIG. 3, the display unit 120 displays the frequency distribution of each node in a different pattern for each class classification result, so that the relationship between the node value and the class classification can be visually shown. . For example, in the intermediate node h1_2, the value of the node is approximately 0 whether the class classification result is the pred_1 class or the pred_2 class. From this, it is considered difficult to classify the pred_1 class and the pred_2 class by using the value of the intermediate node h1_2.

On the other hand, in the intermediate node h1_2, the node value when the class classification result is pred_0 is a value larger than 2, which is different from the node value (approximately 0) for other classes. From this, it is considered possible to classify the pred_0 class and other classes by using the value of the intermediate node h1_2.

Further, in FIG. 3, among the routes from the input node x_3 to the intermediate node h1_0 in the second layer of the hidden layer, a route that passes through the intermediate node h0_4 in the first layer of the hidden layer is shown. This path corresponds to an example of a path from the input node to the final node of the hidden layer.
Regarding this path, the calculation unit 192 calculates the correlation between the distribution of values at input node x_3 and the distribution of values at intermediate node h0_4, and the correlation between the distribution of values at intermediate node h0_4 and the distribution of values at intermediate node h1_0. Calculate the correlation. When the calculation unit 192 calculates a correlation coefficient as a correlation, it is expressed as in equation (1), for example.

Here, ρ indicates a correlation coefficient.
In addition, by normalizing the distribution of node values as described above, it can be regarded as a probability distribution. denotes a random variable of value, and σ _X denotes the standard deviation of the random variable X. Further, Y indicates a random variable of the value of a node at a subsequent stage (on the side receiving data input from an edge), and σ _Y indicates a standard deviation of the random variable Y.
Moreover, σ _XY indicates the covariance between the random variables X and Y.

The correlation between the distributions of these two adjacent values indicates the degree of contribution (magnitude of influence) of the value of the preceding node on the value of the subsequent node. When the edge weighting coefficient between two nodes is a positive value and there is a strong positive correlation, and when the edge weighting coefficient between two nodes is a negative value and there is a strong negative correlation. In either case, it can be said that the contribution of the value of the preceding node to the value of the subsequent node is large. On the other hand, when the correlation between the distributions of the values of two nodes is weak, it can be said that the degree of contribution of the value of the preceding node to the value of the subsequent node is small.

For example, when a person such as a user interprets a target model, it is conceivable that a person, such as a user, attaches importance to edges between nodes with strong correlation, and ignores or downplays edges between nodes with weak correlation. Furthermore, when display unit 120 displays the target model as in the example of FIG. 2, edges between nodes with strong correlation may be displayed darkly, and edges between nodes with weak correlation may be displayed lightly.
Alternatively, the information acquisition device 100 or a person (for example, a user) may simplify the model by deleting edges between nodes with weak correlation among the edges of the target model.

The calculation unit 192 also calculates the product of the correlations of the distribution of node values between each node on the path from the input node to the final node of the hidden layer. For example, in the case of a route from input node x_3 to intermediate node h1_0 via intermediate node h0_4 in FIG. The product of the correlation between the distribution of values of intermediate node h0_4 and the distribution of values of intermediate node h1_0 is calculated.
This product indicates the degree to which data is transmitted along this path. For example, if the correlation can be viewed as a probability (such as through normalization), then this product indicates the probability that the data will follow this path.
The product of correlations calculated by the calculation unit 192 is expressed as in equation (2).

Here, ρ _pr indicates the product of correlations.
M indicates the number of nodes on the route for which the correlation product is calculated. Since the number of nodes is M, the number of nodes between nodes is M-1.
ρ _i indicates the correlation of the distribution of node values between the i-th node and the i+1-th node.

Furthermore, the calculation unit 192 calculates the degree of variation in the values transmitted from the final node of the hidden layer to each output node. When the calculation unit 192 calculates the variance as the degree of dispersion, it is expressed as in equation (3).

Here, X indicates a random variable when the value transmitted from the final node of the hidden layer to each output node (the value obtained by multiplying the output from the final node by the weighting coefficient at the edge) is expressed as a probability distribution. Var(X) indicates the variance of random variable X. E((X-μ) ² ) indicates the expected value of (X-μ) ² .
μ indicates the expected value of the value transmitted from the final node of the hidden layer to each output node, and is expressed as in equation (4).

Here, ρ _pr indicates the product of correlations calculated by equation (2). N indicates the number of output nodes adjacent to the final node of the hidden layer to be calculated. w ₁ , w ₂ , . . . w _N represent weighting coefficients at each edge from the final node of the hidden layer to be calculated to an adjacent output node.

This variance corresponds to an example of an evaluation value of the degree of contribution of the path from the input node of the target model to the final node of the hidden layer to the degree of variation in the value of the output node. The larger the variance, the greater the degree of variation in the values of the output nodes based on the data transmitted through this path. It can be evaluated that the greater the degree of variation in the values of the output nodes, the greater the contribution to the class classification of the input data.

FIG. 4 is a diagram illustrating a display example of a ranking of the degree of contribution of the path from the input node of the target model to the final node of the hidden layer to the degree of variation in the value of the output node, displayed by the display unit 120. FIG. 4 shows an example where the target model is the model shown in FIG. 2.
In the example of FIG. 4, the display unit 120 displays the top 10 routes from the input node to the final node of the hidden layer in descending order of their contribution to the degree of variation in the values of the output nodes. It shows the values of output nodes (expected values) and the average value and standard deviation of the values of these output nodes.

The standard deviation shown in FIG. 4 corresponds to an example of the evaluation value calculated by the calculation unit 192 of the degree of contribution of the path from the input node of the target model to the final node of the hidden layer to the degree of variation in the value of the output node. . It can be evaluated that the larger the value of the standard deviation, the greater the degree of dispersion of the values of the output nodes, and the greater the contribution of the input data to the class classification.

It can be said that this ranking shows a path that is useful for outputting prediction results (class classification of input data) by the target model. For example, when a user interprets a target model, it is expected that the interpretation will become relatively easy by focusing on paths that are useful for outputting prediction results.
Alternatively, when a user tunes a target model, it is expected that tuning can be done relatively efficiently by focusing on tuning the values of weighting coefficients on paths useful for outputting prediction results.

Alternatively, when the user simplifies the target model, the user may delete edges or nodes that have a small contribution to the output of the prediction result, or both of these edges and nodes. An edge or node that is not included in a path useful for outputting a prediction result can be regarded as an edge or node that has a small contribution to outputting a prediction result.
As a result, it is possible to simplify the model by reducing the number of edges or nodes, and it is expected that the influence on the prediction results will be small.

In addition to or in place of the ranking of routes in order of increasing contribution to the degree of variation in output node values, the display unit 120 displays a ranking of routes in descending order of contribution to the degree of variation in output node values. You can do it like this.
When the user deletes edges or nodes that have a small contribution to the output of the prediction result, the user can refer to the ranking of routes in descending order of contribution.

Next, the operation of the information acquisition device 100 will be described with reference to FIG. 5.
FIG. 5 is a flowchart illustrating an example of a processing procedure when the information acquisition device 100 displays the ranking of the degree of contribution of the path from the input node of the target model to the final node of the hidden layer to the degree of variation in the value of the output node. It is.

In the process of FIG. 5, the acquisition unit 191 acquires a plurality of input data, inputs these input data into the target model, and performs model calculations to acquire the distribution of node values at each node ( Step S11). Note that when the distribution of node values is classified into classes and displayed on the display unit 120 as in the example of FIG. The prediction results (class classification results) for are added as metadata.

Next, the calculation unit 192 calculates the correlation of the distribution of node values between adjacent nodes among the input node and the intermediate node (step S12).
Then, for each path from the input node to the final node of the hidden layer, the calculation unit 192 calculates the product of the correlations of the distribution of node values between the nodes on the path (step S13).

Furthermore, the calculation unit 192 calculates, for each path from the input node to the final node of the hidden layer, the value transmitted from the final node of the hidden layer to the output node (the expected value of the output from the final node), the weight of each edge. The variance of the value multiplied by the coefficient is calculated (step S14). This variance corresponds to an example of an evaluation value of the degree of contribution of the path from the input node of the target model to the final node of the hidden layer to the degree of variation in the value of the output node.

Based on the obtained variance, the calculation unit 192 generates a ranking of the degree of contribution of the path from the input node of the target model to the final node of the hidden layer to the degree of variation in the value of the output node, and displays it on the display unit 120. (Step S15).
After step S15, the information acquisition device 100 ends the process of FIG. 5.

As described above, the acquisition unit 191 obtains the distribution of node values for a plurality of input data to the graphical model. The calculation unit 192 calculates the correlation of the distribution of node values between adjacent nodes.
This allows the information acquisition device 100 to indicate the relationship between adjacent nodes. In particular, by showing this correlation, the information acquisition device 100 can determine the degree of contribution (magnitude of influence) of the value of the node in the previous stage (the side that outputs data) to the value of the node in the latter stage (the side that receives data input). can be shown.

When a person (for example, a user) interprets a target model, it is possible to place importance on edges between nodes with strong correlations, and to ignore or downplay edges between nodes with weak correlations. In this way, the information acquisition device 100 can provide information for interpreting the target model.
Furthermore, the information acquisition device 100 or a person (for example, a user) can simplify the model by deleting edges between nodes with weak correlation among the edges of the target model. In this case, it is expected that the impact on the prediction results of the model (classification results of input data) will be relatively small in that edges between nodes with weak correlation are deleted.

The calculation unit 192 also calculates a value indicating the strength of the correlation of the distribution between each node on the path from the input node of the target model to the final node of the hidden layer, and a value that is transmitted from the final node to each output node. An evaluation value of the degree of contribution of the route to the value of the output node is calculated based on a value indicating the degree of variation in values.

The information acquisition device 100 can provide this evaluation value as information for interpreting the target model.
In particular, it can be said that this ranking shows a path that is useful for outputting prediction results (class classification of input data) by the target model. When a person such as a user interprets a target model, it is expected that the interpretation will become relatively easy by focusing on the path that is useful for outputting prediction results.
Alternatively, when a person such as a user tunes a target model, it is expected that tuning can be performed relatively efficiently by focusing on tuning the values of weighting coefficients on paths useful for outputting prediction results.

Alternatively, when a person such as a user simplifies the target model, edges or nodes having a small contribution to the output of the prediction result, or both of these edges and nodes may be deleted. An edge or node that is not included in a path useful for outputting a prediction result can be regarded as an edge or node that has a small contribution to outputting a prediction result.
As a result, it is possible to simplify the model by reducing the number of edges or nodes, and it is expected that the influence on the prediction results will be small.

Further, the calculation unit 192 ranks the routes based on the evaluation value for each route from the input node of the target model to the final node of the hidden layer.
By referring to this ranking result (ranking), users and other people can understand which routes are useful for outputting prediction results by the target model, or which routes are not useful for outputting prediction results by the target model. can.

Furthermore, the display unit 120 displays the distribution of values of a certain node for a plurality of pieces of input data in a different display mode for each class classification result for the input data.
The display unit 120 can visually indicate the relationship between the node value and the class classification by displaying the frequency distribution of each node in a different pattern for each class classification result.

Next, a configuration example according to the embodiment will be described with reference to FIGS. 6 and 7.
FIG. 6 is a diagram illustrating an example of the configuration of the information acquisition device according to the embodiment. The information acquisition device 200 shown in FIG. 6 includes an acquisition section 201 and a calculation section 202.
With this configuration, the acquisition unit 201 obtains the distribution of node values for a plurality of input data to the graphical model. The calculation unit 202 calculates the correlation of the distribution of node values between nodes.

Thereby, the information acquisition device 200 can indicate the relationship between adjacent nodes. In particular, by showing this correlation, the information acquisition device 200 can determine the degree of contribution (magnitude of influence) of the value of the node in the previous stage (the side that outputs data) to the value of the node in the latter stage (the side that receives data input). can be shown.
When a person (for example, a user) interprets a graphical model, it is possible to place importance on edges between nodes with strong correlations, and to ignore or downplay edges between nodes with weak correlations. In this way, the information acquisition device 200 can provide information for interpreting the graphical model.

Furthermore, the information acquisition device 200 or a person (for example, a user) can simplify the model by deleting edges between nodes with weak correlation among the edges of the graphical model. In this case, it is expected that the impact on the prediction results of the model (classification results of input data) will be relatively small in that edges between nodes with weak correlation are deleted.

FIG. 7 is a diagram illustrating an example of a processing procedure in the information acquisition method according to the embodiment.
The information acquisition method shown in FIG. 7 includes the step of calculating the distribution of node values for a plurality of input data to the graphical model (step S21), and the step of calculating the correlation of the distribution of node values between nodes (step S21). S22).

According to the information acquisition method of FIG. 7, the relationship between adjacent nodes can be shown. In particular, by showing this correlation, it is possible to show the degree of contribution (magnitude of influence) of the value of the node in the previous stage (the side that outputs data) with respect to the value of the node in the latter stage (the side that receives data input).
When a person (for example, a user) interprets a graphical model, it is possible to place importance on edges between nodes with strong correlations, and to ignore or downplay edges between nodes with weak correlations. In this way, according to the information acquisition method shown in FIG. 7, information for interpreting the graphical model can be provided.

Furthermore, for example, a person (for example, a user) can simplify the model by deleting edges between nodes with weak correlation among the edges of the graphical model. In this case, it is expected that the impact on the prediction results of the model (classification results of input data) will be relatively small in that edges between nodes with weak correlation are deleted.

FIG. 8 is a schematic block diagram showing the configuration of a computer according to at least one embodiment.
With the configuration shown in FIG. 8, the computer 700 includes a CPU (Central Processing Unit) 710, a main storage device 720, an auxiliary storage device 730, and an interface 740.

Either or both of the information acquisition devices 100 and 200 described above may be implemented in the computer 700. In that case, the operations of each processing section described above are 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 to the main storage device 720, and executes the above processing according to the program. Further, the CPU 710 secures storage areas corresponding to each of the above-mentioned storage units in the main storage device 720 according to the program. Communication between each device and other devices is performed by the interface 740 having a communication function and performing communication under the control of the CPU 710. 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).

When the information acquisition device 100 is installed in the computer 700, the operations of the control unit 190 and each part thereof are 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 to the main storage device 720, and executes the above processing according to the program.
Further, the CPU 710 reserves storage areas corresponding to the storage section 180 and each section thereof in the main storage device 720 according to the program. The communication performed by the communication unit 110 is performed by the interface 740 having a communication function and performing communication under the control of the CPU 710. The functions of the display unit 120 are executed by the interface 740 having a display device and displaying an image on the display screen of the display device under the control of the CPU 710. The function of the operation input unit 130 is performed by the interface 740 having an input device, accepting a user operation, and outputting a signal indicating the accepted user operation to the CPU 710.

When the information acquisition device 200 is installed in the computer 700, the operations of the acquisition unit 201 and the calculation unit 202 are 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 to the main storage device 720, and executes the above processing according to the program.

Note that a program for executing all or part of the processing performed by the control unit 190, the processing performed by the acquisition unit 201, and the processing performed by the calculation unit 202 is recorded on a computer-readable recording medium, and this recording medium Each part may be processed by loading a program recorded in the computer system into the computer system and executing the program. Note that the "computer system" herein includes hardware such as an OS (Operating System) and peripheral devices.
Furthermore, "computer-readable recording media" refers to portable media such as flexible disks, magneto-optical disks, ROM (Read Only Memory), and CD-ROM (Compact Disc Read Only Memory), and hard disks built into computer systems. Refers to storage devices such as Further, the above-mentioned program may be one for realizing a part of the above-mentioned functions, or may be one that can realize the above-mentioned functions in combination with a program already recorded in the computer system.

Although the embodiments of the present invention have been described above in detail with reference to the drawings, the specific configuration is not limited to these embodiments, and includes designs within the scope of the gist of the present invention.

100, 200 information acquisition device 110 communication unit 120 display unit 130 operation input unit 180 storage unit 181 model storage unit 190 control unit 191, 201 acquisition unit 192, 202 calculation unit

Claims (5)

an acquisition unit that calculates the distribution of node values for multiple input data to the graphical model;
The correlation of the distribution between nodes is calculated, and a value indicating the strength of the correlation of the distribution between each node of the path from the input node of the graphical model to the final node of the hidden layer, and each output from the final node. a calculation unit that calculates an evaluation value of the degree of contribution of the route to the value of the output node based on a value indicating the degree of dispersion of the values transmitted to the nodes;
An information acquisition device comprising: The calculation unit ranks the routes based on the evaluation value for each route.
The information acquisition device according to claim 1 . comprising a display unit that displays the distribution of values of a certain node for the plurality of input data in a different display mode for each class classification result for the input data;
The information acquisition device according to claim 1 or claim 2 . a step of calculating the distribution of node values for a plurality of input data to the graphical model; and calculating the correlation of the distribution between the nodes , and calculating the correlation between the nodes of the path from the input node of the graphical model to the final node of the hidden layer. An evaluation value of the degree of contribution of the route to the value of the output node is calculated based on a value indicating the strength of the correlation of the distribution and a value indicating the degree of dispersion of the values transmitted from the final node to each output node. The process of calculating ,
How to obtain information including. determining a distribution of node values for a plurality of input data to the graphical model on a computer;
The correlation of the distribution between nodes is calculated, and a value indicating the strength of the correlation of the distribution between each node of the path from the input node of the graphical model to the final node of the hidden layer, and each output from the final node. calculating an evaluation value of the degree of contribution of the route to the value of the output node based on a value indicating the degree of dispersion of the values transmitted to the node ;
A program to run.

Images