JP7050028B2 - Computer system and machine learning control method - Google Patents

Description

The present invention relates to the control of machine learning.

Even if it is difficult for humans to write down the identification criteria for given data, machine learning technology automatically learns the characteristics of the data and obtains predicted values for unknown data. Is now possible.

In machine learning, a model such as a neural network in deep learning, which has been attracting attention in recent years, is constructed, and parameters that characterize the model are determined using the learning data. In conventional machine learning, initial values of parameters are set based on random numbers and the like, learning processing (iterative calculation) is executed, and the parameters are sequentially updated. At this time, the objective function determined by the parameters is defined, and learning is performed so that the value of the objective function is minimized with respect to the training data.

Japanese Unexamined Patent Publication No. 05-197821

FIG. 9 is a graph showing an example of the relationship between the initial value of the parameter and the accuracy of the output of the model. Here, the horizontal axis represents the parameter and the vertical axis represents the value of the objective function. Further, the objective function is assumed to be a function that outputs the degree of dissociation between the output value of the model and the ideal value. The value F in FIG. 9 represents the value of the objective function that can achieve the desired accuracy.

If the initial values of the parameters are not appropriate, a good model cannot be generated. For example, when P'0 shown in FIG. 9 is set as the initial value of the parameter, the learning result converges to a model (local optimum solution) that is not always optimal. However, when the initial value P'1 is set in the parameter, it converges to the optimum model (global optimum solution). Moreover, since the objective function becomes complicated as the model becomes complicated, it becomes difficult to derive the optimum parameter, that is, the global optimum solution in the complicated model. Therefore, in many cases, the local optimum solution is output as a learning result (final model).

There is no problem if the model corresponding to the locally optimal solution can achieve sufficient accuracy. However, if there is no difference in the accuracy of the model before and after training, or if the desired accuracy is not obtained, machine learning has failed, so it is necessary to give different initial values to the parameters and execute machine learning again. be. Machine learning failures such as those mentioned above are more likely to occur as the model becomes more complex.

The technique described in Patent Document 1 is known for the above-mentioned problems. Patent Document 1 states, "By detecting the amount of change in the sum of squared errors, the correct answer rate, and the amount of change in a neural network that has been learned a predetermined number of times, the learning situation (whether or not it has converged to the optimum solution is determined." After determining whether a local solution has occurred and whether the number of neurons is appropriate), adding neurons to the middle layer, deleting defective neurons, or reinitializing the weights of the defective neurons as necessary. , Re-learn. "

The technique described in Patent Document 1 cannot be applied to models other than neural networks. Further, when the learning data and the model are different, the learning situation, that is, the index for determining the learning failure is also different.

The present invention provides a system and a method that can be applied to various machine learning methods and models and realize efficient machine learning.

A typical example of the invention disclosed in the present application is as follows. That is, it is a computer system that executes machine learning for generating a model that outputs a predicted value with respect to input data, and includes a processor and a storage device connected to the processor, and the machine learning uses the model. It is a process of executing a learning process for updating a defined parameter a plurality of times, and the processor uses at least the number of triggers for defining a determination timing and a machine learning failure sign for generating the model. After acquiring learning determination information that stores detection condition data including detection conditions defined by one evaluation value, setting initial values of the parameters, starting the machine learning, and executing the learning process. The counter representing the number of times the learning process is executed after the initial value of the parameter is set is updated, the detection condition data whose trigger number matches the value of the counter is specified, and the specified detection condition data is specified. Based on the detection conditions included in, it is determined whether or not the machine learning failure sign is detected, and if the machine learning failure sign is detected, the parameters and the counter are initialized, and then the machine. Continue learning.

According to one embodiment of the present invention, it can be applied to various machine learning methods and models, and efficient machine learning can be realized. Issues, configurations and effects other than those described above will be clarified by the description of the following examples.

It is a figure which shows an example of the structure of the computer system of Example 1. FIG. It is a figure which shows an example of the setting screen displayed by the learning part of Example 1. FIG. It is a figure which shows an example of the setting screen displayed by the learning part of Example 1. FIG. It is a graph which shows an example of the relationship between an objective function and a parameter. It is a flowchart explaining the machine learning executed by the computer system of Embodiment 1. FIG. It is a figure which shows an example of the setting screen displayed by the learning part of Example 2. FIG. It is a figure which shows an example of the setting screen displayed by the learning part of Example 2. FIG. It is a flowchart explaining the machine learning executed by the computer system of Example 2. It is a graph which shows an example of the relationship between the initial value of a parameter and the accuracy of the output of a model.

Hereinafter, examples of the present invention will be described with reference to the drawings. However, the present invention is not limited to the description of the examples shown below. It is easily understood by those skilled in the art that a specific configuration thereof can be changed without departing from the idea or purpose of the present invention.

In the configuration of the invention described below, the same or similar configurations or functions are designated by the same reference numerals, and duplicate description will be omitted.

In the present specification, machine learning means a process for generating a model that outputs a predicted value for input data.

Further, in machine learning, learning processing for updating parameters that define a model is executed a plurality of times using training data. The number of learning processes corresponds to the number of epochs in deep learning.

In the first embodiment, a plurality of detection conditions for detecting a failure sign of machine learning are set, and a computer system that performs machine learning detects a failure sign of machine learning based on the plurality of detection conditions. When a sign of machine learning failure is detected, the computer system initializes the parameters and redoes the machine learning. As a result, the success rate of learning can be improved efficiently and quickly. The computer system also provides a means for adjusting the number of learning processes (upper limit) to keep the total calculation time constant in order to prevent an increase in the calculation time due to re-doing machine learning.

(1) System Configuration First, the configuration of the computer system of the first embodiment will be described. FIG. 1 is a diagram showing an example of the configuration of the computer system of the first embodiment.

The computer system is composed of one or more computers 100. FIG. 1 shows a computer system composed of one computer 100 for the sake of simplicity.

The computer 100 has a CPU (Central Processing Unit) 110, a main storage device 111, a sub storage device 112, and an interface 113. Each hardware is connected via an internal bus.

The CPU 110 executes a program stored in the main storage device 111. When the CPU 110 executes processing according to a program, it operates as a functional unit (module) that realizes a specific function. In the following description, when the process is described with the functional unit as the subject, it is shown that the CPU 110 is executing the program that realizes the module.

The main storage device 111 is a memory such as a RAM (Random Access Memory), and stores a program executed by the CPU 110 and data necessary for executing the program.

The main storage device 111 of the first embodiment stores a program that realizes the learning unit 120, and also stores the learning data information 130 and the parameter information 140.

The learning unit 120 executes machine learning for generating a model. The learning unit 120 includes a learning failure detection unit 121, a parameter initialization unit 122, and a calculation time suppression unit 123.

The learning failure detection unit 121 detects the failure sign of machine learning based on the detection condition defined by the evaluation value used for detecting the failure sign of machine learning. The parameter initialization unit 122 sets the initial values of the parameters that define the model. The calculation time suppressing unit 123 suppresses the calculation time required for machine learning.

In addition, each functional unit included in the learning unit 120 may combine a plurality of functional units into one functional unit, or may divide one functional unit into a plurality of functional units for each function.

The training data is data used for machine learning, and is composed of input data to be input to the model and correct answer data. The learning data information 130 is information for associating the input data included in the learning data with the correct answer data, managing the learning data by constructing a mini-batch if necessary, and making it available to the learning unit 120. Parameter information 140 is a parameter that defines a model. The parameter information 140 may store data that defines the structure of the model and the like.

The sub-storage device 112 is a storage device such as an HDD (Hard Disk Drive) and an SSD (Solid State Drive), and permanently stores data.

The interface 113 is an interface for connecting or communicating with an external device. The interface 113 is, for example, an I / O interface, a network interface, or the like. The computer 100 has an interface 113 connected to the external storage device 101 via the network 105, an interface 113 connected to the input / output device 102, and an interface 113 connected to the removable media 103. A terminal having an input / output device 102 may be connected to the computer 100 via the network 105.

The external storage device 101 is a device that provides a storage area for storing data handled by the computer 100. The network 105 connected to the external storage device 101 is, for example, a LAN (Local Area Network), the Internet, or the like. The present invention is not limited to the type of network 105. Further, the connection method of the network 105 may be either wired or wireless.

The input / output device 102 is a device used by the user to operate the computer 100. The input / output device 102 includes a keyboard, a mouse, a touch panel, a display, and the like. The computer 100 may have an input / output device built-in.

The removable media 103 is a storage medium such as a CD and a DVD that can be attached to and detached from the computer 100. The removable media 103 shows an example of means for acquiring the learning data information 130 and the parameter information 140, and is not an essential configuration.

The learning data information 130 and the parameter information 140 stored in the main storage device 111 may be stored in any of the sub storage device 112, the external storage device 101, and the removable media 103. The symbols (A), (B), (C), and (D) in FIG. 1 are for distinguishing information having different storage destinations.

In this case, the CPU 110 reads out the learning data information 130 and the parameter information 140 stored in any of the sub-storage device 112, the external storage device 101, and the removable media 103 when the computer 100 is started up or machine learning is executed. Load into main storage 111. Further, when the power of the computer 100 is turned off, or when the free space of the main storage device 111 is insufficient, the computer 100 may move or copy information from the main storage device 111 to another storage device.

(2) Setting of detection conditions Next, a method of setting the detection conditions set in the computer system of the first embodiment will be described.

The learning unit 120 displays a screen for setting a condition (detection condition) for detecting a failure sign of machine learning on the input / output device 102. 2 and 3 are diagrams showing an example of a setting screen displayed by the learning unit 120 of the first embodiment.

The setting screen 200 shown in FIG. 2 is a screen for setting detection conditions for each of one or more predetermined evaluation values. Details of the evaluation values provided by the present invention will be described later. The setting screen 200 includes an objective function column 201, a correct answer rate column 202, a parameter change rate column 203, and an operation button column 204.

The objective function column 201 is a column for setting a detection condition regarding the value of the objective function, which is one of the evaluation values. The value set in the objective function column 201 is output as detection condition data.

The objective function column 201 includes a trigger count setting column 211, a detection condition setting column 212, and a graph display column 213.

The graph display column 213 is a column for displaying a graph showing the relationship between the value of the target function and the number of learnings. The graph displayed in the graph display field 213 is displayed based on the history of machine learning executed in the past or the result of machine learning currently being executed.

The trigger count setting column 211 is a column for setting the learning count (trigger count) that defines the timing for executing the failure determination process (see FIG. 5) described later based on the detection conditions set in the objective function column 201. be.

The detection condition setting column 212 is a column for setting the detection condition defined by using the objective function. In the detection condition setting field 212, for example, a conditional expression using an objective function and a threshold value is set.

When the detection condition set in the detection condition setting field 212 is satisfied, the learning failure detection unit 121 determines that a failure sign of machine learning has been detected.

The user may set the trigger count and the threshold value by operating the cursor displayed in the graph display column 213 without directly setting the values in the trigger count setting column 211 and the detection condition setting column 212. good.

The correct answer rate column 202 is one of the evaluation values, and is a column for setting a detection condition regarding the correct answer rate of the classification problem in machine learning. Here, the correct answer rate represents the ratio at which the classification result predicted by the model matches the correct answer. The value set in the correct answer rate column 202 is output as detection condition data.

The correct answer rate column 202 includes a trigger number setting column 221, a detection condition setting column 222, and a graph display column 223.

The graph display column 223 is a column for displaying a graph showing the relationship between the correct answer rate and the number of learnings. The graph displayed in the graph display field 223 is displayed based on the history of machine learning executed in the past or the result of machine learning currently being executed.

The trigger count setting column 221 is a column for setting the learning count (trigger count) that defines the timing for executing the failure determination process (see FIG. 5) described later based on the detection conditions set in the correct answer rate column 202. be.

The detection condition setting column 222 is a column for setting a detection condition defined by using the correct answer rate. In the detection condition setting column 222, for example, a conditional expression using the correct answer rate and the threshold value is set.

When the detection condition set in the detection condition setting field 222 is satisfied, the learning failure detection unit 121 determines that a failure sign of machine learning has been detected.

The user may set the trigger count and the threshold value by operating the cursor displayed in the graph display column 223 without directly setting the values in the trigger count setting column 221 and the detection condition setting column 222. good.

The parameter change rate column 203 is a column for setting a detection condition regarding the parameter change rate, which is one of the evaluation values. Here, the parameter change rate represents the degree of change before and after learning the parameters that define the model. The value set in the parameter change rate column 203 is output as detection condition data.

The parameter change rate column 203 includes a trigger number setting column 231, a detection condition setting column 232, and a graph display column 233.

The graph display column 233 is a column for displaying a graph showing the relationship between the parameter change rate and the number of learnings. The graph displayed in the graph display field 233 is displayed based on the history of machine learning executed in the past or the result of machine learning currently being executed.

The trigger count setting column 231 is a column for setting the learning count (trigger count) that defines the timing for executing the failure determination process (see FIG. 5) described later based on the detection conditions set in the parameter change rate column 203. Is.

The detection condition setting column 232 is a column for setting a detection condition defined by using the parameter change rate. In the detection condition setting field 232, for example, a conditional expression using a parameter change rate and a threshold value is set.

When the detection condition set in the detection condition setting field 232 is satisfied, the learning failure detection unit 121 determines that a failure sign of machine learning has been detected.

The user may set the number of triggers and the threshold value by operating the cursor displayed in the graph display field 233 without directly setting the values in the trigger number setting field 231 and the detection condition setting field 232. good.

The operation button column 204 is a column for displaying buttons for performing various operations. The operation button field 204 includes a setting button 241 and a start button 242.

The setting button 241 is a button for setting the learning determination information including the detection condition data set in the objective function column 201, the correct answer rate column 202, and the parameter change rate column 203 in the computer 100. When the setting button 241 is operated, a setting request including learning determination information is output to the computer 100. In this case, the computer 100 stores the learning determination information in the main storage device 111.

The start button 242 is a button for instructing the start of machine learning. When the start button 242 is operated, a start request is output to the computer 100.

The setting screen 300 shown in FIG. 3 is a screen for setting detection conditions for each evaluation value group composed of one or more evaluation values. The setting screen 300 includes a detection condition column 301 and an operation button column 302.

The operation button column 302 is a column for displaying buttons for performing various operations. The operation button column 302 includes an add button 331, a setting button 332, and a start button 333.

The add button 331 is an operation button for adding the detection condition column 301. When the add button 331 is added, the detection condition column 301 is added to the setting screen 300.

The setting button 332 and the start button 333 are the same as the setting button 241 and the start button 242.

The detection condition column 301 includes a trigger number setting column 311, an additional button 312, and a detection condition setting column 313. The value set in the detection condition column 301 is output as the detection condition data.

The trigger number setting column 311 is a column for setting the learning number (trigger number) that defines the timing for executing the failure determination process (see FIG. 5) described later based on the detection condition set in the detection condition column 301. be.

The add button 312 is a button for adding the conditional expression setting field 321 to the detection condition setting field 313. If the add button 312 is operated without the conditional expression setting field 321 in the detection condition setting field 313, the conditional expression setting field 321 is added to the detection condition setting field 313. If the add button 312 is operated while one or more conditional expression setting fields 321 exist in the detection condition setting field 313, the conditional expression setting field 321 and the logical operator setting field 322 are added to the detection condition setting field 313. ..

The conditional expression setting column 321 is a column for setting a conditional expression of evaluation values for determining the appearance of a failure sign of machine learning. The logical operator setting field 322 is a field for setting an application method of a plurality of conditional expressions. In the logical operator setting field 322 of FIG. 3, a pull-down for selecting "AND", "OR", or the like is displayed. By setting a value in the conditional expression setting field 321, it is possible to set a detection condition defined by using at least one evaluation value. It is assumed that the condition specified by AND is indispensable and that any one of the conditions specified by OR is satisfied.

When the detection condition set in the detection condition setting field 313 is satisfied, the learning failure detection unit 121 determines that a failure sign of machine learning has been detected.

Regardless of whether the setting button of the setting screen 200 or 300 is used, the learning determination information including the detection condition data composed of the number of triggers and the detection condition is set in the computer 100.

(3) Evaluation value The computer 100 of the first embodiment detects a failure sign of machine learning by monitoring the behavior of the evaluation value suitable for the model. In the first embodiment, the objective function, the correct answer rate, and the parameter change rate are adopted as evaluation values. The reasons for adopting these evaluation values are as follows.

In the prior art, machine learning is considered to have ended when the change in the objective function is small enough. That is, the objective function is considered to be the most basic evaluation value for determining whether or not learning is stagnant. The determination method is not the value of the objective function itself, but the value of the objective function when the initial value of the parameter is set, or the value of the objective function after the learning process is executed once or a few times. , May be set based on the ratio to the value of the current objective function.

In the case of a model for classifying classes, it is also effective to monitor the accuracy rate of classification results in order to detect signs of machine learning failure. As shown in FIG. 4, the initial value P0 is an initial value in which a model P2 having no practical accuracy is generated, and the initial value P1 is an initial value in which a model P3 having practical accuracy is generated. The value. The value F in FIG. 4 represents the value of the loss function that can achieve the desired accuracy. Regardless of which initial value is set, the value of the objective function is improved by executing the learning process, so that it is difficult to detect the failure sign of machine learning only by monitoring the objective function. On the other hand, the difference in the correct answer rate is remarkable. Therefore, it is considered effective to adopt the correct answer rate as the evaluation value.

Consider the case where the final output of the function f (x) representing the model is output via the ReLU function widely used in deep learning. The ReLU function is given by Eq. (1).

At this time, even if the output of the function f (x) is improved by the learning process, if the value of the function f (x) is 0 or less, the output r (f (x)) remains 0. Therefore, the objective function and the correct answer rate change only when the value of the function f (x) becomes larger than 0. In other words, even if the objective function and the accuracy rate are monitored, the progress of learning may be overlooked. Therefore, the rate of change of the parameter is used as the evaluation value. The rate of change of the parameters can be various, and the following two are shown as examples.

(Example 1) The total value of the parameter values before and after the learning process is calculated, and the ratio of these total values is used as the parameter change rate. Since the parameter value may be negative, it is desirable to calculate the total value of the absolute values of the parameter values.

(Example 2) The ratio of the total value of the change amount of the parameter value before and after the learning process to the total value of the change rate of the parameter value before and after the previous learning process is defined as the parameter change rate. Since the amount of change in the parameter value may be negative, it is desirable to calculate the total value of the absolute values of the amount of change.

(4) Details of processing Next, the processing to be executed when the machine learning start request is received will be described. FIG. 5 is a flowchart illustrating machine learning executed by the computer system of the first embodiment.

In the following description, it is assumed that the upper limit value Nmax of the number of executions of the learning process in machine learning is set in advance.

First, the learning unit 120 initializes the variables N, n, and c_i (step S101).

Specifically, the learning unit 120 sets the variable N to Nmax and n to 0 as initial values. Further, the learning unit 120 sets the variable c_i to the value set in the trigger number setting field 712 corresponding to the i-th evaluation value. Here, the variable N is a variable representing the upper limit of the number of times the learning process being executed at that time should be executed, and the variable n is a variable representing the number of times the learning process is executed after the parameter is initialized. The variable c_i is a variable representing the number of triggers corresponding to the i-th evaluation value.

Next, the parameter initialization unit 122 of the learning unit 120 sets the initial value of the parameter (step S102).

At this time, the parameter initialization unit 122 of the learning unit 120 stores the initial value of the parameter in the parameter information 140. The initial values of the parameters shall be set randomly.

Next, the learning unit 120 executes the learning process using the learning data and updates the parameters (step S103).

At this time, the learning unit 120 stores the parameter update result in the parameter information 140.

Next, the learning unit 120 updates the variable n (step S104).

Specifically, the learning unit 120 sets the variable n as a value obtained by adding 1 to the variable n.

Next, the learning unit 120 determines whether or not the end condition is satisfied (step S105).

Specifically, the learning unit 120 determines that the end condition is satisfied when the variable n is N or more.

When it is determined that the end condition is satisfied, the learning unit 120 ends machine learning.

When it is determined that the end condition is not satisfied, the learning unit 120 determines for each detection condition whether or not there is detection condition data in which the number of triggers c_i included in the detection condition data matches the variable n (step S106). ). Note that i is a subscript that identifies the detection condition data.

When it is determined that the detection condition data whose trigger number c_i included in the detection condition data matches the variable n does not exist, the learning unit 120 returns to step S103 and executes the same process.

When it is determined that the detection condition data whose trigger number c_i included in the detection condition data matches the variable n exists, the learning failure detection unit 121 of the learning unit 120 executes the failure determination process based on the detection condition data. (Step S107).

Specifically, the learning failure detection unit 121 calculates the evaluation value used for the search condition included in the detection condition data. The learning failure detection unit 121 determines whether or not the detection condition set in the detection condition data is satisfied based on the calculated evaluation value.

In the case of the detection condition data set using the setting screen 200, the learning failure detection unit 121 may, for example, determine whether the value of the objective function is smaller than the threshold value, whether the correct answer rate is smaller than the threshold value, or the rate of change. Determine if it is small. The learning failure detection unit 121 determines that the detection condition is satisfied when the value of the objective function is smaller than the threshold value, the correct answer rate is smaller than the threshold value, or the change rate is small.

In the case of the detection condition data set by using the setting screen 300, the learning failure detection unit 121 determines whether or not the detection condition is satisfied based on the combination of the determination results of the conditional expressions of the evaluation values. For example, in the case of the detection condition data corresponding to the "detection condition 2", the learning failure detection unit 121 determines that the detection condition is satisfied when the correct answer rate is smaller than the threshold value or the change rate is smaller than the threshold value.

Next, the learning failure detection unit 121 of the learning unit 120 determines whether or not a failure sign of machine learning has been detected based on the result of the failure determination process (step S108).

Specifically, the learning failure detection unit 121 determines that a machine learning failure sign has been detected if the detection condition is satisfied.

If it is determined that the machine learning failure sign has not been detected, the learning unit 120 returns to step S103 and executes the same process.

When it is determined that the failure sign of machine learning is detected, the parameter initialization unit 122 of the learning unit 120 sets a new initial value in the parameter (step S109).

At this time, the parameter initialization unit 122 of the learning unit 120 stores the initial value of the parameter in the parameter information 140. The initial values of the parameters shall be set randomly.

Next, the calculation time suppressing unit 123 of the learning unit 120 updates the upper limit value N to Nn (step S110). Further, the calculation time suppressing unit 123 of the learning unit 120 sets the variable n to 0 (step S111), then returns to step S103 and executes the same process.

By the above processing, the total number of executions of the learning process in machine learning is controlled so as not to exceed the original upper limit value Nmax. As a result, the calculation time required for machine learning can be suppressed.

(5) Application Example Next, an application example of Example 1 will be described.

(Application example 1) Deep learning is adopted as an algorithm for machine learning. A loss function is used as the objective function.

(Application example 2) An EM algorithm is adopted as a machine learning algorithm. For example, when learning a normal mixture distribution using an EM algorithm, efficient machine learning can be realized by changing the EM algorithm as shown in Example 1. The log-likelihood is usually used as the objective function.

(Application example 3) The learning data is a numerical vector derived from a DNA sequence. Various methods can be considered for converting a DNA sequence into a numerical vector. Here, an example called one hot encoding will be described. In one hot encoding, each base is represented by a sequence of three 0s and one 1. The position of 1 differs depending on the base type. When the arrangement of 0s and 1s is, for example, A = 1000, T = 0100, G = 0010, C = 0001, the array ATGC can be expressed as 10001000100001. By inputting a DNA sequence as a vector consisting of numerical values, it is possible to generate a model that recognizes a sequence having a specific function on the genome, for example, a promoter region.

(6) Features and Effects of Example 1 Next, the features and effects of machine learning executed by the computer system described in Example 1 will be described.

The computer system of the first embodiment detects a failure sign of machine learning based on a plurality of detection conditions defined by at least one evaluation value. By using multiple detection conditions, it is possible to accurately detect signs of failure in machine learning that generate various models.

In addition, the timing for determining the occurrence of a failure sign of machine learning can be set for each detection condition. By reducing the timing, it is possible to detect signs of machine learning failure at an early stage. Further, by changing the timing of the failure determination processing for each detection condition, it is possible to improve the detection accuracy of the failure sign and speed up the processing by reducing the processing load.

Moreover, since re-learning is performed only by initializing numerical parameters, it can be applied to machine learning methods and models other than neural networks.

Further, by adjusting the upper limit value (variable N) of the number of iterations of the learning process in consideration of the number of executions of the learning process so far, it is possible to suppress an increase in the calculation time required for machine learning.

The second embodiment is different in that the detection condition is set in consideration of the upper limit value N of the number of learnings. For example, when the upper limit value is larger than a certain value, the detection accuracy of the failure sign of machine learning can be improved by lengthening the determination timing. Further, when the upper limit value is smaller than a certain value, by shortening the determination timing, it is possible to secure the number of learnings while detecting the failure sign of machine learning at an early stage. Hereinafter, Example 2 will be described with a focus on the differences from Example 1.

The configuration of the computer system of the second embodiment is the same as that of the first embodiment. The hardware configuration and software configuration of the computer 100 of the second embodiment are the same as those of the first embodiment. In the second embodiment, the detection condition data to be set is different.

6 and 7 are diagrams showing an example of a setting screen displayed by the learning unit 120 of the second embodiment.

The setting screen 600 shown in FIG. 6 is a screen for setting detection conditions for each evaluation value defined in advance. The setting screen 600 includes an objective function column 601, a correct answer rate column 602, a parameter change rate column 603, and an operation button column 604.

The objective function column 601 is a column for setting a detection condition regarding the value of the objective function, which is one of the evaluation values.

The objective function column 601 includes an add button 611 and a detection condition setting table 612. The objective function column 601 may be provided with a column for displaying a graph as in the first embodiment.

The add button 611 is a button for adding an entry to the detection condition setting table 612.

The detection condition setting table 612 is a table for setting detection conditions. The detection condition setting table 612 includes an entry composed of the application condition 621, the number of triggers 622, and the detection condition 623. One entry corresponds to one detection condition data.

The application condition 621 is a field for storing the condition for the upper limit value N in order to select the search condition to be applied. As will be described later, in the second embodiment, the failure determination process is executed using the detection condition satisfying the application condition 621.

The trigger count 622 is a field for storing the learning count (trigger count) that defines the timing for executing the failure determination process based on the detection condition set in the objective function column 601.

The detection condition 623 is a field for storing the detection condition defined by using the objective function. In the detection condition 623, for example, a conditional expression using an objective function and a threshold value is stored.

In the second embodiment, the detection conditions defined by using the objective function are set for each range of the upper limit value.

The correct answer rate column 602 is a column for setting a detection condition regarding the correct answer rate, which is one of the evaluation values.

The correct answer rate column 602 includes an additional button 631 and a detection condition setting table 632. The correct answer rate column 602 may be provided with a column for displaying a graph as in the first embodiment.

The add button 631 is a button for adding an entry to the detection condition setting table 632.

The detection condition setting table 632 is a table for setting the detection condition data. The detection condition setting table 632 includes an entry composed of the application condition 641, the number of triggers 642, and the detection condition 643. One entry corresponds to one detection condition data.

The application condition 641 is a field for storing the condition for the upper limit value N in order to select the detection condition to be applied. As will be described later, in the second embodiment, the failure determination process is executed using the detection condition satisfying the application condition 641.

The trigger number 642 is a field for storing the learning number (trigger number) that defines the timing for executing the failure determination process based on the detection condition set in the correct answer rate column 602.

The detection condition 643 is a field for storing the detection condition defined by using the correct answer rate. In the detection condition 643, for example, a conditional expression using the correct answer rate and the threshold value is stored.

In the second embodiment, the detection condition defined by using the correct answer rate is set for each range of the upper limit value.

The parameter change rate column 603 is a column for setting a detection condition regarding the parameter change rate, which is one of the evaluation values.

The parameter change rate column 603 includes an additional button 651 and a detection condition setting table 652. The parameter change rate column 603 may be provided with a column for displaying a graph as in the first embodiment.

The add button 651 is a button for adding an entry to the detection condition setting table 652.

The detection condition setting table 652 is a table for setting the detection condition data. The detection condition setting table 652 includes an entry composed of the application condition 661, the number of triggers 662, and the detection condition 663. One entry corresponds to one detection condition data.

The application condition 661 is a field for storing the condition for the upper limit value N in order to select the detection condition to be applied. As will be described later, in the second embodiment, the failure determination process is executed using the detection condition satisfying the application condition 661.

The trigger number 662 is a field for storing the learning number (trigger number) that defines the timing for executing the failure determination process based on the detection condition set in the parameter change rate column 603.

The detection condition 663 is a field for storing the detection condition defined by using the parameter change rate. In the detection condition 643, for example, a conditional expression using a parameter change rate and a threshold value is stored.

In the second embodiment, the detection condition defined by using the parameter change rate is set for each range of the upper limit value.

Since the operation button column 604 is the same column as the operation button column 204, the description thereof will be omitted.

The setting screen 700 shown in FIG. 7 is a screen for setting detection conditions for each evaluation value group. The setting screen 700 includes a detection condition column 701 and an operation button column 702.

The addition button 731, the setting button 732, and the start button 733 of the operation button column 702 are the same as the addition button 331, the setting button 332, and the start button 333 of the operation button column 302.

The detection condition column 701 includes an application condition setting column 711, a trigger number setting column 712, an add button 713, and a detection condition setting column 714. The value set in the detection condition column 701 is output as the detection condition data.

The applicable condition setting column 711 is a column for setting a condition for the upper limit value in order to select a detection condition. As will be described later, in the second embodiment, the failure determination process is executed using the detection condition satisfying the application condition setting field 711.

Since the trigger count setting field 712 and the add button 713 are the same as the trigger count setting field 311 and the add button 312, the description thereof will be omitted. Further, the conditional expression setting column 721 and the logical operator setting column 722 of the detection condition setting column 714 are the same as the conditional expression setting column 321 and the logical operator setting column 322 of the detection condition setting column 313. Omit.

Regardless of which of the setting screens 200 and 300 is used, the learning determination information including the detection condition data including the application condition, the number of triggers, and the detection condition is set in the computer 100.

FIG. 8 is a flowchart illustrating machine learning executed by the computer system of the second embodiment.

In the following description, it is assumed that the upper limit value Nmax of the number of executions of the learning process in machine learning is set in advance.

After the processes of steps S101 and S102 are executed, the learning unit 120 selects the detection condition data to be applied from the learning determination information (step S151).

Specifically, the learning unit 120 selects detection condition data satisfying the applicable conditions based on the current upper limit value N.

The processing of step S101, step S102, step S103 to step S106, and step S108 to step S111 is the same as that of the first embodiment. However, the value of c_i referred to in step S106 shall be set in steps S101 and S110 based on the value of N at the time of setting on the setting screen 600 or 700 and parameter initialization.

In step S107 of the second embodiment, the learning failure detection unit 121 executes the failure determination process using the detection condition data selected in step S151. Since the specific processing is the same as that of the first embodiment, the description thereof will be omitted.

The computer system of the second embodiment can adjust the determination timing of the learning failure sign according to the upper limit value.

The present invention is not limited to the above-described embodiment, and includes various modifications. Further, for example, the above-described embodiment describes the configuration in detail in order to explain the present invention in an easy-to-understand manner, and is not necessarily limited to the one including all the described configurations. Further, it is possible to add, delete, or replace a part of the configuration of each embodiment with other configurations.

Further, each of the above configurations, functions, processing units, processing means and the like may be realized by hardware by designing a part or all of them by, for example, an integrated circuit. The present invention can also be realized by a software program code that realizes the functions of the examples. In this case, a storage medium in which the program code is recorded is provided to the computer, and the processor included in the computer reads out the program code stored in the storage medium. In this case, the program code itself read from the storage medium realizes the function of the above-described embodiment, and the program code itself and the storage medium storing it constitute the present invention. Examples of the storage medium for supplying such a program code include a flexible disk, a CD-ROM, a DVD-ROM, a hard disk, an SSD (Solid State Drive), a CD-R, a DVD-R, a magnetic tape, and a non-volatile material. Memory card, ROM, etc. are used.

In addition, the program code that realizes the functions described in this embodiment can be implemented in a wide range of programs or script languages such as assembler, C / C ++, Perl, Python, and Java (registered trademark).

Further, by distributing the program code of the software that realizes the functions of the embodiment via the network, the program code is stored in a storage means such as a hard disk or a memory of a computer or a storage medium such as a CD-RW or a CD-R. The processor included in the computer may read and execute the program code stored in the storage means or the storage medium.

In the above-described embodiment, the control lines and information lines show what is considered necessary for explanation, and do not necessarily indicate all the control lines and information lines in the product. All configurations may be interconnected.

100 Computer 101 External storage device 102 Input / output device 103 Removable media 105 Network 110 CPU
111 Main storage device 112 Secondary storage device 113 Interface 120 Learning unit 121 Learning failure detection unit 122 Parameter initialization unit 123 Calculation time suppression unit 130 Learning data information 140 Parameter information 200, 300, 600, 700 Setting screen

Claims (8)

A computer system that performs machine learning to generate a model that outputs predicted values for input data.
A processor and a storage device connected to the processor are provided.
The machine learning is a process of executing a learning process of updating the parameters defining the model a plurality of times.
The processor
Learning judgment information that stores detection condition data including detection conditions defined by the number of triggers that specify the judgment timing and at least one evaluation value used to detect the failure sign of machine learning for generating the model. Acquired,
After setting the initial values of the parameters, the machine learning is started.
After executing the learning process, the counter indicating the number of times the learning process is executed after the initial value of the parameter is set is updated.
The detection condition data whose trigger count matches the value of the counter is specified, and the detection condition data is specified.
Based on the detection conditions included in the specified detection condition data, it is determined whether or not the failure sign of the machine learning is detected.
A computer system comprising initializing the parameters and the counter and then continuing the machine learning when a sign of failure of the machine learning is detected. The computer system according to claim 1.
The processor
It manages the upper limit of the number of times the learning process is executed in the machine learning,
When the value of the counter is equal to or greater than the upper limit value, the machine learning is terminated and the machine learning is terminated.
A computer system characterized in that when a sign of failure of machine learning is detected, a value obtained by subtracting the upper limit value by the value of the counter is set as a new upper limit value, and then the counter is initialized. The computer system according to claim 2.
The detection condition data includes an application condition for determining whether or not to apply the detection condition data, which is defined by using the upper limit value.
The processor
Based on the upper limit value, the detection condition data satisfying the applicable condition is selected, and the detection condition data is selected.
A computer system characterized in that the detection condition data whose trigger count matches the value of the counter is specified from the selected detection condition data. The computer system according to claim 1.
A computer system characterized in that the detection condition is a conditional expression using the evaluation value and the threshold value. A machine learning control method executed by a computer system to generate a model that outputs predicted values for input data.
The computer system has a processor and a computer having a storage device connected to the processor.
The machine learning is a process of executing a learning process of updating the parameters defining the model a plurality of times.
The machine learning control method is
The computer stores detection condition data including a detection condition defined by a number of triggers defining a determination timing and at least one evaluation value used to detect a sign of failure of machine learning for generating the model. The first step to acquire learning judgment information and
A second step in which the computer sets initial values for the parameters and initiates machine learning.
A third step in which the computer updates a counter representing the number of times the learning process is executed after the initial value of the parameter is set after the computer executes the learning process.
A fourth step in which the computer identifies the detection condition data whose trigger count matches the value of the counter.
A fifth step in which the computer determines whether or not the machine learning failure sign has been detected based on the detection conditions included in the specified detection condition data.
Control of machine learning, characterized in that the computer comprises a sixth step of initializing the parameters and the counter and then continuing the machine learning when a sign of failure of the machine learning is detected. Method. The machine learning control method according to claim 5.
The computer system manages an upper limit of the number of times the learning process is executed in the machine learning.
The method for controlling machine learning includes a step in which the computer ends the machine learning when the value of the counter is equal to or greater than the upper limit value.
The sixth step is
When the computer detects a sign of failure in machine learning, the step of setting a value obtained by subtracting the upper limit value by the value of the counter as a new upper limit value, and
A method for controlling machine learning, wherein the computer comprises a step of initializing the counter. The machine learning control method according to claim 6.
The detection condition data includes an application condition for determining whether or not to apply the detection condition data, which is defined by using the upper limit value.
The fourth step is
A step in which the computer selects the detection condition data satisfying the applicable condition based on the upper limit value.
A method for controlling machine learning, wherein the computer includes a step of specifying the detection condition data whose trigger count matches the value of the counter from the selected detection condition data. The machine learning control method according to claim 5.
The detection condition is a control method for machine learning, characterized in that it is a conditional expression using the evaluation value and the threshold value.

Images