JP6901423B2 - Information processing equipment, information processing terminals, and programs - Google Patents

Description

The present invention relates to an information processing device, an information processing terminal, and a program, and more particularly to a technique for adjusting the correct answer rate of a neural network.

In recent years, speed-up of CPU (Central Processing Unit) and GPU (Graphics Processing Unit), increase in memory capacity, and machine learning technology using a neural network have been rapidly advancing. For this reason, machine learning using learning data on the order of hundreds of thousands to one million has become possible, and highly accurate identification technology and classification technology are being established (see Non-Patent Document 1).

Yangqing Jia, Evan Shelhamer, Jeff Donahue, Sergey Karayev, Jonathan Long, Ross Girshick, Sergio Guadarrama, and Trevor Darrell. Caffe: Convolutional architecture for fast feature embedding. In Proceedings of the 22nd ACM international conference on Multimedia (pp. 675-678) ). ACM.

A large amount of computational cost is required to perform machine learning based on a large amount of learning data. In addition, a huge amount of labor is required for preparing a large amount of learning data and for preprocessing to use the prepared learning data for machine learning. When the model creator who generated the learning model wants to make a profit by letting a third party use the learning model, it is convenient to be able to change the performance of the learning model according to the consideration charged to the user. ..

The present invention has been made in view of these points, and an object of the present invention is to provide a learning model with different correct answer rates.

The first aspect of the present invention is an information processing device. This device includes a machine learning unit that generates model parameters of a neural network for detecting a target task, a parameter modification unit that generates a modification parameter obtained by modifying a part of the model parameter, and a test input for the modification parameter. It is provided with a correct answer rate calculation unit that calculates the correct answer rate of data, and a storage unit that stores the correct answer rate calculated by the correct answer rate calculation unit in association with each modification pattern of the model parameter to be modified by the parameter modification unit. ..

The parameter modification unit replaces a selection unit that selects a layer for modifying the model parameter among a plurality of layers constituting the neural network and a model parameter of the layer selected by the selection unit with a predetermined value. The replacement unit to be replaced, the comparison unit for comparing the correct answer rate after replacement with the target correct answer rate for modification of the model parameter, and (1) when the correct answer rate is higher than the target correct answer rate, the said The selection unit reselects the layer on the input layer side of the selected layer, and (2) when the correct answer rate is lower than the target correct answer rate, the selection unit selects the layer on the output layer side of the selected layer. Reselection may be provided, and (3) a modification control unit that outputs the model parameter replaced by the replacement unit as a modification parameter when the difference between the correct answer rate and the target correct answer rate falls within a predetermined range may be provided. ..

The information processing device includes a permission correct answer rate acquisition unit that acquires a correct answer rate permitted to the user of the information processing terminal from an information processing terminal different from the information processing device, and a correct answer rate stored in the storage unit. Among them, the modification pattern selection unit that selects the modification pattern corresponding to the correct answer rate similar to the correct answer rate acquired by the permission correct answer rate acquisition unit, and the model parameter modified based on the selected modification pattern. A model parameter transmission unit that transmits to the information processing terminal via a communication network may be further provided.

The information processing device includes a task data acquisition unit that acquires task data input to the neural network from an information processing terminal different from the information processing device via a communication network, and a parameter modification unit that has not been modified. Even if the detection execution unit that acquires the result of inputting the task data into the model parameter and the result transmission unit that transmits the result acquired by the detection execution unit to the information processing terminal via the communication network are further provided. Good.

A second aspect of the present invention is an information processing terminal. This terminal is a model parameter related to the neural network generated to detect the target task, the first model parameter having the higher correct answer rate of the target task, and the target task than the first model parameter. A model parameter acquisition unit that acquires two model parameters, a second model parameter with a low correct answer rate, and a usage limit number acquisition that acquires an upper limit of the number of times the first model parameter is used among the two model parameters. A detection execution unit that acquires the result of inputting task data to be detected in one of the model parameters of the first model parameter and the second model parameter, and the first model parameter by the detection execution unit. The first model parameter is selected when the number of times of use is less than the upper limit value, and a model parameter selection unit is provided which selects the second model parameter when the number of times of use reaches the upper limit value. Here, the detection execution unit acquires the output result of the task data by using the model parameter selected by the model parameter selection unit.

A third aspect of the present invention is a program. This program has a function of causing a computer to generate a model parameter of a neural network for detecting a target task, a function of generating a modified parameter obtained by modifying a part of the model parameter, and test input data in the modified parameter. The function of calculating the correct answer rate of the above and the function of associating the calculated correct answer rate with each modification pattern of the model parameter to be modified and storing it in the storage unit are realized.

A fourth aspect of the present invention is also a program. This program is a model parameter related to the neural network generated for detecting the target task on the computer, and is the first model parameter having the higher correct answer rate of the target task, and the first model parameter is higher than the first model parameter. A function to acquire two model parameters, a second model parameter having a low correct answer rate of the target task, a function to acquire an upper limit of the number of times the first model parameter is used among the two model parameters, and the above. A function and selection of selecting the first model parameter when the cumulative number of times of use of the first model parameter is less than the upper limit value, and selecting the second model parameter when the cumulative number of times of use reaches the upper limit value. A function to acquire the output result of task data is realized by using the obtained model parameters.

According to the present invention, it is possible to provide learning models with different correct answer rates.

It is a figure which shows typically the structure of the neural network which concerns on embodiment. It is a figure which shows typically the functional structure of the information processing apparatus which concerns on embodiment. It is a figure which shows typically the internal structure of the parameter change part which concerns on embodiment. It is a figure which shows typically the data structure of the modification pattern database which concerns on embodiment. It is a figure which shows typically the whole structure of the information processing system which concerns on embodiment. It is a flowchart for demonstrating the flow of information processing executed by the information processing apparatus which concerns on embodiment. It is a figure which shows typically the functional structure of the information processing terminal which concerns on a modification.

<Outline of the embodiment>
FIG. 1 is a diagram schematically showing a configuration of a neural network according to an embodiment. Hereinafter, an outline of the embodiment will be described with reference to FIG.

The neural network used by the information processing apparatus according to the embodiment has a general configuration including an input layer, an intermediate layer, and an output layer as a whole. As shown in the upper part of FIG. 1, in the information processing apparatus according to the embodiment, the basic learning data input to the input layer propagates through the intermediate layer, and finally the data string output by the output layer and the basic learning data. The error from the correct label corresponding to the data is calculated using the loss function. The information processing apparatus updates the weight set in the intermediate layer by using the error back propagation method based on the calculated error.

A part of the intermediate layer of the neural network used by the information processing apparatus according to the embodiment is modified. Specifically, as shown in the lower part of FIG. 1, some nodes are removed in the modification target layer, which is a layer selected from the intermediate layers. In the lower part of FIG. 1, white circles are normal nodes in the neural network, and black circles are the removed nodes.

Neural networks are designed to achieve final identification accuracy by propagating the output of each node in the middle layer to adjacent layers. Therefore, removing the nodes breaks the connection between the nodes, which lowers the correct answer rate of the learning model.

Note that removing a node is equivalent to modifying the coefficient to be multiplied when the output of that node is input to the next node, that is, the value of the model parameter of the learning model to 0. Eliminating nodes also reduces the complexity of model parameters and is often used to compress training models. For example, when executing a learning model on a device having scarce computing resources such as an IoT (Internet Of Things) device, a learning model in which nodes are removed to reduce the size may be used.

The information processing apparatus according to the embodiment calculates the correct answer rate of the learning model while changing the modification pattern applied to the model parameters. The information processing device holds the modification pattern applied to the model parameter and the correct answer rate in association with each other. As a result, the information processing device 1 according to the embodiment can provide learning models with different correct answer rates.

<Functional configuration of information processing device 1>
FIG. 2 is a diagram schematically showing a functional configuration of the information processing device 1 according to the embodiment. The information processing device 1 includes a storage unit 10 and a control unit 11.

The storage unit 10 includes a ROM (Read Only Memory) for storing the BIOS (Basic Input Output System) of the computer that realizes the information processing device 1, a RAM (Random Access Memory) that serves as a work area for the information processing device 1, and an OS (OS). An operating system), an application program, and a large-capacity storage device such as an HDD (Hard Disk Drive) or SSD (Solid State Drive) that stores various information referred to when the application program is executed.

The control unit 11 is a processor such as a CPU or GPU of the information processing device 1, and by executing a program stored in the storage unit 10, the machine learning unit 110, the parameter modification unit 111, the correct answer rate calculation unit 112, and the permitted correct answer are executed. It functions as a rate acquisition unit 113, a modification pattern selection unit 114, a model parameter transmission unit 115, a task data acquisition unit 116, a detection execution unit 117, and a result transmission unit 118.

Note that FIG. 2 shows an example in which the information processing device 1 is composed of a single device. However, the information processing device 1 may be realized by computing resources such as a plurality of processors and memories, such as a cloud computing system. In this case, each unit constituting the control unit 11 is realized by executing a program by at least one of a plurality of different processors.

The machine learning unit 110 generates model parameters of the neural network for detecting the target task. The machine learning unit 110 stores the generated model parameters in the storage unit 10.

The parameter modification unit 111 generates a modification parameter obtained by modifying a part of the model parameter generated by the machine learning unit 110. Specifically, the parameter modification unit 111 generates a new model parameter obtained by replacing a part of the weighting coefficients constituting the model parameter of the neural network with another value as the modification parameter. The value in which the parameter modification unit 111 replaces the weighting coefficient may be any value as long as it differs from the original model parameter, but is 0 as an example.

The correct answer rate calculation unit 112 calculates the correct answer rate of the test input data in the modified parameters generated by the parameter modification unit 111. Here, the "test input data" is a fixed data set prepared for measuring the correct answer rate of the model parameter. Since the correct answer rate calculation unit 112 calculates the correct answer rate of the modified parameters using the same test input data, the correct answer rates of different modified parameters can be calculated under the same conditions.

The storage unit 10 stores the correct answer rate calculated by the correct answer rate calculation unit 112 in association with each modification pattern of the model parameter to be modified by the parameter modification unit 111. As a result, the information processing device 1 according to the embodiment can provide learning models with different correct answer rates.

FIG. 3 is a diagram schematically showing the internal configuration of the parameter modification unit 111 according to the embodiment. The parameter modification unit 111 according to the embodiment includes a selection unit 1110, a replacement unit 1111, a comparison unit 1112, and a modification control unit 1113.

The selection unit 1110 selects a layer for modifying model parameters from a plurality of layers constituting the neural network. The replacement unit 1111 replaces the model parameter of the layer selected by the selection unit 1110 with a predetermined value. The comparison unit 1112 compares the correct answer rate after replacement by the replacement unit 1111 with the target correct answer rate for which the model parameters are modified.

When the correct answer rate is higher than the target correct answer rate, the modification control unit 1113 causes the selection unit 1110 to reselect a layer on the input layer side of the selected layer. Generally, in a neural network, it is known that a layer closer to the input layer has a greater influence on the correct answer rate than a layer far from the input layer. The modification control unit 1113 can reduce the correct answer rate after replacement by the replacement unit 1111 by bringing the modification target layer to be modified by the parameter closer to the input layer.

When the correct answer rate is lower than the target correct answer rate, the modification control unit 1113 causes the selection unit 1110 to reselect a layer on the output layer side of the selected layer. For the same reason as described above, the modification control unit 1113 can increase the correct answer rate after replacement by the replacement unit 1111 by moving the modification target layer away from the input layer (that is, moving it to the output layer side).

Further, when the difference between the correct answer rate and the target correct answer rate falls within a predetermined range, the modification control unit 1113 outputs the model parameter replaced by the replacement unit 1111 as the modification parameter. Here, the "predetermined range" is a "target correct answer rate error permissible range" that the parameter modification unit 111 refers to when determining an error tolerance range from the target correct answer rate of the parameter modification. The specific value in the predetermined range may be determined by an experiment in consideration of the difficulty of adjusting the correct answer rate, and is, for example, a range of 2% before and after the target correct answer rate.

By repeating the above operation for a plurality of target correct answer rates, the parameter modification unit 111 can generate a modification pattern for each of a plurality of types of correct answer rates.

FIG. 4 is a diagram schematically showing the data structure of the modified pattern database according to the embodiment. The modification pattern database is held by the storage unit 10 and managed by the parameter modification unit 111.

As shown in FIG. 4, the modification pattern database stores the replacement location of the parameter, the value of the parameter before replacement, and the percentage of the model size after replacement with respect to the model size before replacement for each of the plurality of correct answer rates. doing. For example, in FIG. 4, in the model having a correct answer rate of 90%, the fourth node of the fifth layer and the eighth node of the fifth layer are at least replaced. The model size after replacement is 70% of the model size before replacement, which is 30% compressed.

Returning to the description of FIG. The information processing device 1 according to the embodiment can allow a user who is a third party to use the generated learning model. At this time, the information processing device 1 changes the correct answer rate of the learning model that can be used by the user according to the consideration paid by the user.

Therefore, the permission correct answer rate acquisition unit 113 acquires the correct answer rate permitted by the information processing terminal 2 connected to the information processing device 1 via the communication network N. Specifically, the permission correct answer rate acquisition unit 113 stores a terminal identifier for identifying the information processing terminal 2 and a correct answer rate permitted for the information processing terminal 2 in association with each other (not shown). To obtain the correct answer rate permitted for the information processing terminal 2. Alternatively, the permission correct answer rate acquisition unit 113 may receive and acquire the correct answer rate stored in the internal storage of the information processing terminal 2 connected to the information processing device 1 from the information processing terminal 2.

The modification pattern selection unit 114 selects a modification pattern corresponding to the correct answer rate similar to the correct answer rate acquired by the permission correct answer rate acquisition unit 113 among the correct answer rates stored in the storage unit 10. Here, the "correct answer rate similar to the correct answer rate acquired by the permission correct answer rate acquisition unit 113" is the correct answer rate equal to or higher than the correct answer rate acquired by the permission correct answer rate acquisition unit 113 among the correct answer rates stored in the storage unit 10. It means the correct answer rate closest to the correct answer rate acquired by the permission correct answer rate acquisition unit 113.

The model parameter transmission unit 115 transmits the model parameters modified based on the modification pattern selected by the modification pattern selection unit 114 to the information processing terminal 2 via the communication network N. As a result, the information processing terminal 2 can execute the learning model on its own terminal using the acquired model parameters and realize the target task. In other words, once the information processing terminal 2 has acquired the model parameters from the information processing device 1, the information processing terminal 2 can execute the target task offline even if it is not connected to the information processing device 1.

As described above, the model parameter in which the correct answer rate is set to be low tends to be smaller in size than the model parameter in which the correct answer rate is set to be high. Therefore, even when the calculation resource of the information processing terminal 2 is small, it may be possible to execute the model parameter if the correct answer rate is set low. It is also meaningful to set the correct answer rate of the model parameters low in the sense that it can be executed even with a device with few calculation resources.

On the other hand, if the device can communicate with the information processing device 1 via the communication network N even if the device has low calculation resources, the correct answer rate can be increased by having the information processing device 1 take over part of the calculation of the model parameters. The target task can also be achieved by using high model parameters (that is, model parameters that have a large computational load at run time).

Therefore, the information processing device 1 includes a task data acquisition unit 116 that acquires task data to be input to model parameters of a neural network from an information processing terminal 2 different from the information processing device 1 via a communication network N. The detection execution unit 117 acquires the result of inputting the task data acquired by the task data acquisition unit 116 into the model parameters that have not been modified by the parameter modification unit 111. The result transmission unit 118 transmits the result acquired by the detection execution unit 117 to the information processing terminal 2 via the communication network N. As a result, the information processing terminal 2 executes the target task using a model parameter having a high correct answer rate, which has a large calculation load at the time of execution, under the condition that the information processing terminal 2 can communicate with the information processing device 1 using the communication network N. be able to.

FIG. 5 is a diagram schematically showing the overall configuration of the information processing system S according to the embodiment. As shown in FIG. 5, the information processing system S according to the embodiment includes an information processing device 1 and an information processing terminal 2 that are connected to each other via a communication network N in a manner in which they can communicate with each other. The information processing device 1 provides the information processing terminal 2 with a learning model in which a part of the intermediate layer is modified to reduce the correct answer rate. In this case, when the information processing terminal 2 executes the learning model in the local environment, the learning model with a reduced correct answer rate is executed.

The information processing terminal 2 receives an input API (Application Program Interface) and an output API from the information processing device 1. By using the input API, the information processing terminal 2 can transmit the data to be processed to the information processing device 1 via the communication network N. Further, the information processing terminal 2 can receive the execution result of the learning model from the information processing device 1 via the communication network N by using the output API. As a result, the information processing terminal 2 selects between an online high-performance learning model and an offline low-load learning model according to the correct answer rate required for the target task and the communication environment in which the information processing terminal 2 is placed. it can.

<Processing flow of information processing executed by the information processing device 1 according to the embodiment>
FIG. 6 is a flowchart for explaining the flow of information processing executed by the information processing apparatus 1 according to the embodiment. The process in this flowchart starts, for example, when the information processing device 1 is activated.

The machine learning unit 110 generates a model parameter of the neural network for detecting the target task (S2). The parameter modification unit 111 generates a modification parameter obtained by modifying a part of the model parameter generated by the machine learning unit 110 (S4). The correct answer rate calculation unit 112 calculates the correct answer rate of the test input data in the modified parameter modified by the parameter modification unit 111 (S6). The storage unit 10 stores the correct answer rate calculated by the correct answer rate calculation unit 112 in association with each modification pattern of the model parameter to be modified by the parameter modification unit 111 (S8).

<Effects of the information processing device 1 according to the embodiment>
As described above, according to the information processing apparatus 1 according to the embodiment, it is possible to provide learning models with different correct answer rates.

Although the present invention has been described above using the embodiments, the technical scope of the present invention is not limited to the scope described in the above embodiments, and various modifications and changes can be made within the scope of the gist thereof. is there. For example, the specific embodiment of the distribution / integration of the device is not limited to the above embodiment, and all or a part thereof may be functionally or physically distributed / integrated in any unit. Can be done. Also included in the embodiments of the present invention are new embodiments resulting from any combination of the plurality of embodiments. The effect of the new embodiment produced by the combination has the effect of the original embodiment together.

<Modification example>
In the above, the information processing terminal 2 has described the case where the learning model having a high correct answer rate is executed online and the learning model having a low correct answer rate is executed offline. However, for example, when the information processing terminal 2 has sufficient computing resources and the information processing terminal 2 alone can execute a learning model having a high correct answer rate and a large calculation load, the information processing apparatus 1 is fully loaded on the information processing terminal 2. A size learning model may be provided.

On the other hand, if the information processing device 1 permits the information processing terminal 2 to execute the full-size learning model without limitation, the administrator of the information processing device 1 can obtain the benefit of using the learning model by a third party. It can be damaged.

Therefore, the information processing device 1 may impose restrictions on the execution of the full-size learning model by the information processing terminal 2. Hereinafter, the execution restrictions of the full-size learning model imposed on the information processing terminal 2 will be described in more detail.

FIG. 7 is a diagram schematically showing the functional configuration of the information processing terminal 2 according to the modified example. The information processing terminal 2 according to the modified example includes a storage unit 20 and a control unit 21.

The storage unit 20 stores a ROM that stores the BIOS and the like of the computer that realizes the information processing terminal 2, a RAM that is a work area of the information processing device 1, an OS and an application program, and various information that is referred to when the application program is executed. It is a large-capacity storage device such as an HDD or SSD for storing.

The control unit 21 is a processor such as a CPU or GPU of the information processing terminal 2, and by executing a program stored in the storage unit 20, the model parameter acquisition unit 210, the usage limit number acquisition unit 211, the detection execution unit 212, And functions as a model parameter selection unit 213.

The model parameter acquisition unit 210 is a model parameter related to the neural network generated to detect the target task, and is the first model parameter having the higher correct answer rate of the target task and the target task than the first model parameter. Two model parameters, a second model parameter with a low correct answer rate and a second model parameter, are acquired. The model parameter acquisition unit 210 acquires, for example, as the first model parameter, a model parameter that has not been modified by the parameter modification unit 111.

The usage limit number acquisition unit 211 acquires the upper limit value of the number of times of use of the first model parameter from the information processing device 1 among the two model parameters acquired by the model parameter acquisition unit 210. The upper limit of the number of times the first model parameter is used is set according to, for example, the consideration paid by the user of the information processing terminal 2 to the information processing device 1.

The detection execution unit 212 acquires the result of inputting the task data to be detected into one of the model parameters of the first model parameter and the second model parameter. Specifically, the detection execution unit 212 acquires the output result of the task data by using the model parameter selected by the model parameter selection unit 213.

The model parameter selection unit 213 selects the first model parameter when the number of times the first model parameter is used by the detection execution unit 212 is less than the upper limit value acquired from the information processing device 1. Further, the model parameter selection unit 213 selects the second model parameter when the number of times the first model parameter is used by the detection execution unit 212 reaches the upper limit value. As a result, the information processing terminal 2 can execute the target task using the model parameters having a high correct answer rate until the upper limit of the number of times of use permitted by the information processing device 1 is reached.

1 ... Information processing device 10 ... Storage unit 11 ... Control unit 110 ... Machine learning unit 111 ... Parameter modification unit 1110 ... Selection unit 1111 ... Replacement unit 1112 ... Comparison Unit 1113 ... Modification control unit 112 ... Correct answer rate calculation unit 113 ... Allowed correct answer rate acquisition unit 114 ... Modification pattern selection unit 115 ... Model parameter transmission unit 116 ... Task data acquisition unit 117・ ・ ・ Detection execution unit 118 ・ ・ ・ Result transmission unit 2 ・ ・ ・ Information processing terminal 20 ・ ・ ・ Storage unit 21 ・ ・ ・ Control unit 210 ・ ・ ・ Model parameter acquisition unit 211 ・ ・ ・ Usage limit number acquisition unit 212 ... Detection execution unit 213 ... Model parameter selection unit S ... Information processing system N ... Communication network

Claims (7)

A machine learning unit that generates neural network model parameters to detect the target task,
A parameter modification unit that generates a modification parameter obtained by modifying a part of the model parameter,
A correct answer rate calculation unit that calculates the correct answer rate of the test input data in the modified parameters, and
A storage unit that stores the correct answer rate calculated by the correct answer rate calculation unit in association with each modification pattern of the model parameter to be modified by the parameter modification unit.
Information processing device equipped with. The parameter modification part
A selection unit for selecting a layer for modifying the model parameters from a plurality of layers constituting the neural network, and a selection unit.
A replacement unit that replaces the model parameter of the layer selected by the selection unit with a predetermined value,
A comparison unit that compares the correct answer rate after replacement with the target correct answer rate that is the target of modification of the model parameters.
(1) When the correct answer rate is higher than the target correct answer rate, the selection unit is made to reselect a layer on the input layer side of the selected layer, and (2) when the correct answer rate is lower than the target correct answer rate. , The selection unit reselects a layer on the output layer side of the selected layer, and (3) a model in which the replacement unit is replaced when the difference between the correct answer rate and the target correct answer rate falls within a predetermined range. A modification control unit that outputs parameters as modification parameters,
The information processing apparatus according to claim 1. The permission correct answer rate acquisition unit that acquires the correct answer rate permitted to the user of the information processing terminal from an information processing terminal different from the information processing device, and
Among the correct answer rates stored in the storage unit, a modified pattern selection unit that selects a modified pattern corresponding to a correct answer rate similar to the correct answer rate acquired by the permitted correct answer rate acquisition unit, and
A model parameter transmission unit that transmits the model parameters modified based on the selected modification pattern to the information processing terminal via a communication network.
The information processing apparatus according to claim 1 or 2. A task data acquisition unit that acquires task data to be input to the neural network from an information processing terminal different from the information processing device via a communication network.
A detection execution unit that acquires the result of inputting the task data into the model parameter that has not been modified by the parameter modification unit, and a detection execution unit.
A result transmission unit that transmits the result acquired by the detection execution unit to the information processing terminal via the communication network, and a result transmission unit.
The information processing apparatus according to any one of claims 1 to 3, further comprising. The first model parameter, which is a model parameter related to the neural network generated to detect the target task and has a higher correct answer rate of the target task, and the correct answer rate of the target task is lower than that of the first model parameter. A model parameter acquisition unit that acquires two model parameters, a second model parameter, and
Of the two model parameters, a usage limit number acquisition unit that acquires an upper limit of the number of times the first model parameter is used, and a usage limit number acquisition unit.
A detection execution unit that acquires the result of inputting task data to be detected in one of the model parameters of the first model parameter and the second model parameter.
A model parameter that selects the first model parameter when the number of times the first model parameter is used by the detection execution unit is less than the upper limit value, and selects the second model parameter when the number of times of use reaches the upper limit value. With a selection part,
The detection execution unit acquires the output result of task data using the model parameters selected by the model parameter selection unit.
Information processing terminal. On the computer
A function to generate model parameters of the neural network to detect the target task, and
A function to generate a modified parameter obtained by modifying a part of the model parameter, and
A function to calculate the correct answer rate of the test input data in the modified parameters, and
A function to associate the calculated correct answer rate for each modification pattern of the model parameter to be modified and store it in the storage unit.
A program that realizes. On the computer
The first model parameter, which is a model parameter related to the neural network generated to detect the target task and has a higher correct answer rate of the target task, and the correct answer rate of the target task is lower than that of the first model parameter. The function to acquire the second model parameter and the two model parameters,
Of the two model parameters, the function of acquiring the upper limit of the number of times the first model parameter is used, and
A function of selecting the first model parameter when the cumulative number of times of use of the first model parameter is less than the upper limit value, and selecting the second model parameter when the cumulative number of times of use reaches the upper limit value.
A function to acquire the output result of task data using the selected model parameters,
A program that realizes.

Images