Learning device, estimation device, learning method, and program
Patent Information
- Application Number
- JP2024511041
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2024-09-25
- Publication Date
- 2024-12-04
- Estimated Expiration
- 2042-03-31
AI Technical Summary
Setting hyperparameter values for machine learning models is challenging, as existing methods require searching a wide range of values and are difficult to optimize effectively.
A learning device and method that uses an evaluation function to determine the estimation accuracy of a model, updating the model to bring the output value of the evaluation function closer to a reference value, thereby simplifying the setting of hyperparameter values and preventing overfitting.
This approach allows for easier setting of hyperparameter values, improving model accuracy and robustness against changes in parameter values, while reducing the risk of overfitting and shortening the learning time without requiring extensive computational resources.
Abstract
Description
Learning device, estimation device, learning method, and recording medium
[0001] The present invention relates to a learning device, an estimation device, a learning method, and a recording medium.
[0002] One method for setting hyperparameters for a machine learning model involves performing learning using multiple hyperparameter values and selecting one of the multiple hyperparameter values based on the learning results. For example, a parameter adjustment device described in Patent Literature 1 sets multiple parameter values for one of two hyperparameters, hyperparameter A, and sets a fixed value for the other hyperparameter, hyperparameter B. The parameter adjustment device transmits combinations of each value of hyperparameter A and a fixed value of hyperparameter B to a learning device, obtains the accuracy rate for each combination, and approximates the relationship between the value of hyperparameter A and the accuracy rate using a function. The parameter adjustment device also approximates the relationship between the value of hyperparameter A and the accuracy rate for other values of hyperparameter B using a function, and finds the combination of the value of hyperparameter A and the value of hyperparameter B that results in the highest accuracy rate.
[0003] Japanese Patent Application Publication No. 2018-15992
[0004] When training a model, it is possible to set hyperparameters to allow for a wide range of possible parameter values to be explored, but it is difficult to set the hyperparameter values appropriately.
[0005] An example of a purpose of this disclosure is to provide a learning device, an estimation device, a learning method, and a recording medium that can solve the above-mentioned problems.
[0006] According to a first aspect of the present invention, a learning device includes a standard setting means for setting a standard value for the degree of estimation accuracy to be achieved in further machine learning based on an evaluation function value indicating an evaluation of the estimation accuracy of a model generated in machine learning, and a learning means for performing machine learning to update the model so as to bring the output value of the evaluation function closer to the standard value.
[0007] According to a second aspect of the present invention, an estimation device includes an estimation unit that calculates an estimation value for an estimation target using the model that has been updated through machine learning so as to bring the output value of the evaluation function closer to a reference value that indicates the extent to which estimation accuracy of the model should be achieved through further machine learning, the reference value being set based on an evaluation function value that indicates an evaluation of the estimation accuracy of the model generated through machine learning.
[0008] According to a third aspect of the present invention, a learning method includes a computer setting a reference value for the degree of estimation accuracy to be achieved in further machine learning based on an evaluation function value indicating an evaluation of the estimation accuracy of a model generated in machine learning, and learning the model so as to bring the output value of the evaluation function closer to the reference value.
[0009] According to a fourth aspect of the present invention, the recording medium is a recording medium that records a program for causing a computer to execute the following steps: setting a reference value for the degree of estimation accuracy to be achieved in further machine learning based on an evaluation function value that indicates an evaluation of the estimation accuracy of a model generated in machine learning; and learning the model so as to bring the output value of the evaluation function closer to the reference value.
[0010] According to the present invention, it is possible to set the values of hyperparameters for adjusting the breadth of the search for parameter values relatively easily.
[0011] FIG. 1 is a diagram illustrating an example of the configuration of a learning device according to an embodiment; FIG. 2 is a diagram illustrating an example of division of learning data by a data acquisition unit according to an embodiment; FIG. 3 is a diagram illustrating an example of robustness against changes in parameter values; FIG. 4 is a diagram illustrating a first example of a processing procedure in which a learning device according to an embodiment learns a model; FIG. 5 is a diagram illustrating an example of a start condition and an end condition for learning using a reference value according to an embodiment; FIG. 6 is a diagram illustrating a second example of a processing procedure in which a learning device according to an embodiment learns a model; FIG. 7 is a diagram illustrating a third example of a processing procedure in which a learning device according to an embodiment learns a model; FIG. 8 is a diagram illustrating an example of the configuration of an estimation device according to an embodiment; FIG. 9 is a diagram illustrating another example of the configuration of a learning device according to an embodiment; FIG. 10 is a diagram illustrating an example of processing procedures in a learning method according to an embodiment; and FIG. 11 is a schematic block diagram illustrating the configuration of a computer according to at least one embodiment.
[0012] Embodiments of the present invention will be described below, but the following embodiments do not limit the scope of the invention. Furthermore, not all of the combinations of features described in the embodiments are necessarily essential to the solution of the invention. FIG. 1 is a diagram showing an example of the configuration of a learning device according to an embodiment. In the configuration shown in FIG. 1, the learning device 100 includes a communication unit 110, a display unit 120, an operation input unit 130, a memory unit 180, and a control unit 190. The memory unit 180 includes a model memory unit 181. The control unit 190 includes a data acquisition unit 191, a reference setting unit 192, and a learning unit 193.
[0013] The learning device 100 learns a model. The learning device 100 may be configured using a computer such as a personal computer (PC) or a workstation. The communication unit 110 communicates with other devices. For example, the communication unit 110 may communicate with a device that stores learning data and receive the learning data.
[0014] The display unit 120 has a display screen such as a liquid crystal panel or an LED (Light Emitting Diode) panel, and displays various images. The operation input unit 130 has input devices such as a keyboard and a mouse, and accepts user operations.
[0015] For example, for hyperparameters whose values are set by the user, the display unit 120 may display a hyperparameter value input screen, and the operation input unit 130 may accept a user operation to input the hyperparameter value.
[0016] The storage unit 180 stores various data. The storage unit 180 is configured using a storage device included in the learning device 100. The model storage unit 181 stores a model to be learned. However, the model to be learned by the learning device 100 is not limited to the one stored in the model storage unit 181. For example, the model to be learned by the learning device 100 may be implemented using hardware and configured as a device separate from the learning device 100. In this case, the storage unit 180 may be configured without including the model storage unit 181.
[0017] The control unit 190 performs various processes by controlling each unit of the learning device 100. The functions of the control unit 190 may be performed, for example, by a CPU (Central Processing Unit) included in the learning device 100 reading and executing a program from the storage unit 180.
[0018] The data acquisition unit 191 acquires learning data. For example, when the communication unit 110 receives learning data from another device, the data acquisition unit 191 extracts the learning data from the data received by the communication unit 110. The data acquisition unit 191 also divides the acquired learning data into training data, confirmation data, and test data. The following describes an example in which the data acquisition unit 191 acquires supervised sample data as learning data. The supervised sample data here can be a set whose elements are combinations of inputs to a model and correct answers that are the output of the model for those inputs.
[0019] 2 is a diagram showing an example of division of learning data by the data acquisition unit 191. In the example of Fig. 2, the data acquisition unit 191 divides the supervised sample data acquired as learning data into training data, confirmation data, and test data. Specifically, the data acquisition unit 191 divides multiple elements, which are combinations of inputs to a model and correct answers that are output from the model in response to those inputs, included in the supervised sample data into training data elements, confirmation data elements, and test data elements.
[0020] Training data is used to adjust the parameter values of a model. A parameter here refers to a variable included in the model whose value is updated by a learning algorithm. For example, when training a perceptron-type neural network using backpropagation, the weight coefficients between nodes and the bias at each node are examples of parameters.
[0021] The confirmation data is used to adjust the values of hyperparameters in model training. Hyperparameters here refer to parameters related to model training, such as parameters for setting the behavior of the learning algorithm, other than parameters whose values are subject to update by the learning algorithm. For example, when training a perceptron-type neural network using backpropagation, the learning rate is an example of a parameter. Furthermore, when the structure of a neural network is variable, values related to the neural network structure, such as the number of hidden layers and the number of nodes per layer, can also be treated as hyperparameters.
[0022] For example, the learning device 100 may set multiple values for the hyperparameters and use the confirmation data to select one of the hyperparameter values. In this case, the learning device 100 adjusts the parameter values for each of the set hyperparameter values using training data, and then evaluates the model using the confirmation data. The learning device 100 may then select the hyperparameter value with the highest evaluation. Note that the reference value of the evaluation function value, which will be described later, can also be treated as a hyperparameter. However, as will be described later, the learning device 100 sets only one value for the reference value of the evaluation function value at a time.
[0023] The test data is used as data for evaluating a model obtained by training. For example, the training device 100 may evaluate the model obtained by training using the test data. Then, based on the evaluation results, the training device 100 or a user may decide whether to adopt the obtained model or to retrain the model.
[0024] Training data and validation data can be said to be data used to build a model. Test data can be said to be data used to evaluate a model. Training data can be said to be data used to update parameter values of a model. Validation data and test data can be said to be data other than data used to update parameter values of a model.
[0025] However, the division of the supervised sample data performed by the data acquisition unit 191 is not limited to division into training data, confirmation data, and test data. For example, when a model is tested through trial operation in an actual usage environment, it is not necessary to obtain test data from the supervised sample data. In this case, the data acquisition unit 191 may divide the acquired supervised sample data into training data and confirmation data.
[0026] The standard setting unit 192 sets a standard value for the evaluation function value for the evaluation function used in model training. The standard setting unit 192 is an example of standard setting means. The evaluation function here is a function that indicates an evaluation of the estimation accuracy of the model. The estimation accuracy of the model indicated by the evaluation function can also be said to be the degree of fit of the model to the data input to the model. A high evaluation indicated by the evaluation function can also be said to mean a high degree of fit of the model to the input data.
[0027] Therefore, the evaluation function can also be said to be a function that indicates the degree of fit of the model to the data input to the model. For example, when training data is input to the model, the evaluation function value can be said to indicate the degree of fit of the model to the training data.
[0028] In the following, an example will be described in which the learning device 100 uses, as an evaluation function, a function in which the smaller the value, the higher the evaluation, such as an error function or a cross entropy loss function. In this case, the smaller the value of the evaluation function, the better the fit of the model to the input data.
[0029] In the following, an example will be described in which the minimum value of the evaluation function value is 0. The error here may be a value indicated by "1 - precision". Alternatively, the loss or error here may be a value based on the distance between the output value of the model and the correct value, such as L1 loss or L2 loss. However, the evaluation function used by the learning unit 193 is not limited to a specific one. For example, the learning unit 193 may use, as the evaluation function, a function that indicates a higher evaluation as the evaluation function value increases.
[0030] The reference value here is a value that represents the standard for the degree of estimation accuracy to be achieved for training data in model learning. The reference value can also be said to be a value that specifies the degree of fit of the model to the training data. In other words, the reference value can also be said to represent a standard for preventing overfitting. The closer the reference value is to 0, the more the model that is generated will fit the training data.
[0031] The learning unit 193 learns the model. In particular, the learning unit 193 learns the model so as to bring the evaluation function value as close as possible to the reference value. The learning unit 193 corresponds to an example of a learning means.
[0032] The standard setting unit 192 sets the evaluation function value indicating that the degree of fit of the model to the training data is more limited than the maximum degree of fit as the standard value. A smaller evaluation function value indicates a higher degree of fit, and when the minimum evaluation function value is 0, the standard setting unit 192 sets a value greater than 0 as the standard value of the evaluation function value. The learning unit 193 learns the model so as to bring the evaluation function value as close as possible to the standard value.
[0033] The reference value set by the reference setting unit 192 is used as a hyperparameter for adjusting the breadth of the search for parameter values. When the reference value is relatively small (i.e., close to 0), the learning unit 193 is considered to be relatively unlikely to search for another local solution after reaching a local solution in which the evaluation function value is the reference value or a value close to it in the search for parameter values. On the other hand, when the reference value is relatively large, the learning unit 193 is considered to be relatively likely to search for another local solution after reaching a local solution in which the evaluation function value is the reference value or a value close to it in the search for parameter values. In order for the learning unit 193 to search for another local solution after reaching a local solution in the search for parameter values, a solution search method that stochastically determines the next search point, such as a stochastic gradient method, may be used.
[0034] It is expected that the learning unit 193 can prevent overfitting by learning a model using a reference value of the evaluation function value. By preventing overfitting, for example, it is expected that the learning unit 193 can obtain parameter values that are relatively robust against changes in the parameter values. Here, being robust against changes in the parameter values means that the evaluation function value does not increase significantly even if the parameter values change slightly.
[0035] FIG. 3 is a diagram showing an example of robustness against changes in parameter values. The horizontal axis of the graph in FIG. 3 represents the parameter value, and the vertical axis represents the evaluation function value. FIG. 3 shows an example in which the evaluation function value is minimized when the value of the parameter w is w1. Line L11 shows an example in which the parameter value w1 is relatively robust against changes in the parameter value. On the other hand, line L12 shows an example in which the parameter value w1 is relatively not robust against changes in the parameter value.
[0036] Comparing the cases of line L11 and line L12, the increase in the evaluation function value is smaller in the case of line L11 when the value of the parameter w changes slightly. From this, it is expected that when the learning unit 193 obtains a local solution for the parameter values indicated by line L11, it will be easier to obtain parameter values that are suitable for new data when learning progresses and new data is presented, compared to when the learning unit 193 obtains a local solution for the parameter values indicated by line L12. In this way, it is expected that parameter values that are robust to changes in parameter values will result in higher model accuracy as learning progresses than parameter values that are not robust to changes in parameter values.
[0037] It is expected that overfitting can be prevented by the learning unit 193 learning a model using the reference value of the evaluation function value. On the other hand, if the reference value is too large, it may be impossible to obtain a model with a high evaluation indicated by the evaluation function value, that is, the accuracy of the model obtained by learning may be low.
[0038] Regarding the setting of the reference value, the reference setting unit 192 sets the reference value based on the evaluation function values obtained in the learning up to that point. For example, the reference setting unit 192 may set the evaluation function value obtained by applying the confirmation data to the model in a past epoch as the reference value for the new epoch. The evaluation function value obtained by applying data such as the confirmation data to the model may be the average value of the evaluation function values obtained for each element included in the data.
[0039] The term "epoch" or "one epoch" used here refers to one round of model learning repeatedly performed by the learning unit 193 using the same supervised sample data. The number of times an epoch is repeated is also referred to as the "epoch count." An epoch corresponds to an example of one unit of model learning repeatedly performed by the learning unit 193.
[0040] When the learning unit 193 uses a loss function as an evaluation function, the evaluation function value obtained by applying training data to the model is also referred to as training loss, the evaluation function value obtained by applying verification data to the model is also referred to as verification loss, and the evaluation function value obtained by applying test data to the model is also referred to as test loss.
[0041] When the learning unit 193 uses an error function as the evaluation function, the evaluation function value obtained by applying the training data to the model is also referred to as the training error, the evaluation function value obtained by applying the verification data to the model is also referred to as the verification error, and the evaluation function value obtained by applying the test data to the model is also referred to as the test error.
[0042] The reference setting unit 192 may set any one of the training loss, validation loss, test loss, training error, validation error, or test error as the reference value of the evaluation function. Alternatively, the reference setting unit 192 may set a value obtained by calculation based on any one of the training loss, validation loss, test loss, training error, validation error, or test error, such as a value obtained by multiplying the training loss by a predetermined coefficient, as the reference value of the evaluation function. Alternatively, the reference setting unit 192 may set a value obtained by calculation based on a combination of two or more indices of the training loss, validation loss, test loss, training error, validation error, or test error, such as the average value of the training loss and the validation loss, as the reference value of the evaluation function.
[0043] The standard setting unit 192 may set an evaluation function including the standard value, thereby setting the standard value. For example, the standard setting unit 192 may set an evaluation function J * (g) may be set.
[0044]
[0045] g indicates the model to be learned. J(g) indicates the original evaluation function (the evaluation function without the reference value). b indicates the reference value. "||" indicates the absolute value. If J(g) ≥ b, then J * (g)=J(g). In this case, the learning unit 193 uses the evaluation function J * In the learning using (g), the parameter values are searched for in the same way as when J(g) is used as the evaluation function.
[0046] On the other hand, if J(g)<b, then J * (g) = 2b - J(g). In this case, the sign of the term "-J(g)" is minus (-), and the evaluation function J * J(g) has the opposite gradient to J(g). When J(g) is used as an evaluation function in a solution search algorithm based on the gradient method such as backpropagation, the learning unit 193 searches for parameter values so that the value of the evaluation function J(g) approaches the minimum value 0 as much as possible. * When (g) is used, the learning unit 193 uses the evaluation function J * The parameter value is searched for so that the value of (g) approaches the reference value b as closely as possible.
[0047] Evaluation function J * In the domain of the evaluation function J(g), in the part where the output value of the evaluation function J(g) is equal to the reference value b or is greater than the reference value b, the evaluation function J(g) outputs the same value as the output value of the evaluation function J(g). * In the domain of the evaluation function J(g), in the portion where the output value of the evaluation function J(g) is smaller than the reference value b, the evaluation function J(g) outputs a value larger than the output value of the evaluation function J(g). * (g) is also called a restricted evaluation function.
[0048] The standard setting unit 192 may set the evaluation function value of an epoch in which the evaluation function value satisfies a predetermined standard, among the epochs already executed by the learning unit 193, as the standard value for the next epoch. In this case, the evaluation function value referred to for selecting the epoch and the evaluation function value set as the standard value may be evaluation function values of different data.
[0049] For example, the reference setting unit 192 may select an epoch whose confirmation error satisfies a predetermined criterion. In this case, the predetermined criterion may be that the confirmation error for each epoch is ranked within a predetermined order of highest value when sorted in ascending order. The reference setting unit 192 may set the reference value based on the training error for the selected epoch. Alternatively, the reference setting unit 192 may set the reference value to a value obtained by calculation using the training error, such as a value obtained by adding a predetermined value to the training error or a value obtained by subtracting a predetermined value from the training error.
[0050] For example, the reference setting unit 192 may set the evaluation function value of the epoch with the smallest evaluation function value among the epochs already executed by the learning unit 193 as the reference value for the next epoch. In this case, the evaluation function value referred to for selecting the epoch and the evaluation function value set as the reference value may be evaluation function values of different data.
[0051] Furthermore, for example, the reference setting unit 192 may select an epoch with the smallest training error from among the epochs already executed by the learning unit 193, and set the validation error for that epoch as the reference value for the next epoch. In this case, the reference setting unit 192 can set the reference value by referring to good learning results, in that it selects an epoch with the smallest training error. Also, since validation errors are generally considered to be larger than training errors, the reference setting unit 192 sets a relatively large reference value. By having the learning unit 193 learn the model based on this reference value, it is expected that overfitting will be less likely to occur.
[0052] Alternatively, the reference setting unit 192 may select an epoch with the smallest confirmation error from among the epochs already executed by the learning unit 193, and set the training error in that epoch as the reference value for the next epoch. In an epoch with a small confirmation error, the generalization performance of the obtained model is expected to be relatively high. By setting the training error in the epoch with the smallest confirmation error as the reference value, the reference setting unit 192 sets the training error when a model with relatively high generalization performance is obtained as the reference value. By having the learning unit 193 learn the model based on this reference value, it is expected that it will be easier to search for a solution that will result in a model with relatively high generalization performance, and in this respect, it is expected that overfitting will be avoided.
[0053] The learning unit 193 may be configured to have the reference setting unit 192 determine whether to update the reference value each time it performs learning for one epoch. Figure 4 is a diagram showing a first example of a processing procedure in which the learning device 100 performs model learning. In the processing of Figure 4, the learning unit 193 performs learning for the first epoch (step S101). In learning for the first epoch, the learning unit 193 has not yet performed any epochs, and the reference setting unit 192 has not set a reference value. Therefore, the learning unit 193 performs model learning without setting a reference value. If a reference value is not set, the learning unit 193 performs model learning so as to bring the evaluation function value closer to the minimum value.
[0054] When the learning unit 193 finishes learning for the first epoch, the reference setting unit 192 sets a reference value for the evaluation function value based on the learning result for the first epoch (step S102). The learning unit 193 performs learning for one epoch using the reference value set by the reference setting unit 192 (step S103).
[0055] Next, the reference setting unit 192 determines whether the index value in the epoch most recently executed by the learning unit 193 is the smallest value among the index values in the epochs that the learning unit 193 has already executed (step S104). As described above, the index value here may be any of the training loss, validation loss, test loss, training error, validation error, or test error. Furthermore, the index value used by the reference setting unit 192 for the determination in step S104 may be different from the index value used to set the reference value.
[0056] If the learning unit 193 determines that the index value in the most recently executed epoch is the smallest value among the index values in the epochs that the learning unit 193 has already executed (step S104: YES), the learning unit 193 updates the reference value to the value obtained from the learning result in the most recent epoch (step S111).
[0057] Next, the learning unit 193 determines whether a predetermined learning termination condition is met (step S112). The learning termination condition here is a condition for determining whether the learning unit 193 should terminate model learning. The learning termination condition here is not limited to a specific condition. For example, the learning termination condition may be a condition that the learning unit 193 has completed learning for a predetermined number of epochs. Alternatively, the learning termination condition may be a condition that the error of the obtained model is equal to or less than a predetermined error threshold.
[0058] If the learning unit 193 determines that the learning end condition is not met (step S112: NO), the process returns to step S103. On the other hand, if the learning unit 193 determines that the learning end condition is met (step S112: YES), the learning device 100 ends the process of FIG.
[0059] On the other hand, in step S104, if the standard setting unit 192 determines that the index value in the epoch most recently executed by the learning unit 193 is not the smallest value among the index values in the epochs that the learning unit 193 has already executed (step S104: NO), the processing proceeds to step S112.
[0060] It is conceivable that the performance of the model will not be stable in the early stages of model learning. Therefore, the learning unit 193 may perform learning using a reference value after learning has progressed to a certain extent. For example, as will be described later, until a predetermined reference value use start condition is met, the learning unit 193 may perform model learning so as to bring the evaluation function value closer to 0 without using a reference value. Then, after the reference value use start condition is met, the learning unit 193 may perform model learning so as to bring the evaluation function value closer to the reference value. The reference value use start condition here is a condition for determining when the learning unit 193 will start learning using a reference value, or a condition for determining whether the learning unit 193 will perform learning using a reference value.
[0061] Furthermore, the learning unit 193 may perform learning without using a reference value in the final stage of model learning to improve the performance of the model. For example, the learning unit 193 may perform learning for the last 100 epochs without using a reference value. For example, as described below, the learning unit 193 may perform model learning so as to bring the evaluation function value closer to the reference value until a predetermined reference value use termination condition is met. After the reference value use termination condition is met, the learning unit 193 may perform model learning so as to bring the evaluation function value closer to 0 without using the reference value. The reference value use termination condition here is a condition for determining when the learning unit 193 will end learning using a reference value, or a condition for determining whether the learning unit 193 will end learning using a reference value.
[0062] FIG. 5 is a diagram illustrating examples of start and end conditions for learning using a reference value. The horizontal axis of the graph in FIG. 5 represents the number of epochs. The vertical axis represents the error. Line L21 represents an example of the training error. Line L22 represents an example of the validation error. When the training error becomes equal to or less than the threshold Et, the reference setting unit 192 may set the validation error in the epoch in which the training error is smallest as the reference value. The learning unit 193 may perform learning without setting a reference value in the epoch before the reference setting unit 192 sets the reference value, and may perform learning based on the reference value in the epoch in which the reference setting unit 192 sets the reference value. Furthermore, the learning unit 193 may perform learning without setting a reference value after the number of epochs reaches M, and terminate learning when the number of epochs reaches N. Here, M and N are both positive integers, and M<N.
[0063] 6 is a diagram showing a second example of a processing procedure in which the learning device 100 learns a model. In the processing of FIG. 6, the learning unit 193 performs learning for one epoch without setting a reference value (step S201). Next, the reference setting unit 192 determines whether a predetermined reference value use start condition is satisfied (step S202). The reference value use start condition here is not limited to a specific condition. For example, as in the example of FIG. 5, the reference value use start condition may be a condition that the training error is equal to or less than a predetermined threshold, but is not limited to this.
[0064] If the reference setting unit 192 determines that the reference value use start condition is not satisfied (step S202: NO), the process returns to step S201. On the other hand, if the reference value use start condition is satisfied (step S202: YES), the reference setting unit 192 sets a reference value (step S211). Specifically, the reference setting unit 192 selects an epoch with the smallest index value from among the epochs already executed by the learning unit 193, and sets a reference value based on the learning result of that epoch. As described above, the index value here may be any of training loss, validation loss, test loss, training error, validation error, and test error. Furthermore, the index value used by the reference setting unit 192 to select an epoch may be different from the index value used to set the reference value. The learning unit 193 performs learning for one epoch based on the reference value set by the reference setting unit 192 (step S212).
[0065] Next, the reference setting unit 192 determines whether the index value in the epoch most recently executed by the learning unit 193 is the smallest value among the index values in the epochs that the learning unit 193 has already executed (step S213). As described above, the index value here may be any of the training loss, validation loss, test loss, training error, validation error, or test error. Furthermore, the index value used by the reference setting unit 192 for the determination in step S213 may be different from the index value used to set the reference value.
[0066] If the learning unit 193 determines that the index value in the most recently executed epoch is the smallest value among the index values in the epochs that the learning unit 193 has already executed (step S213: YES), the learning unit 193 updates the reference value to the value obtained from the learning result in the most recent epoch (step S221).
[0067] Next, the learning unit 193 determines whether a predetermined condition for ending the use of the reference value is satisfied (step S222). The condition for ending the use of the reference value is not limited to a specific condition. For example, as shown in the example of FIG. 5, the condition for ending the use of the reference value may be, but is not limited to, a condition that the learning unit 193 has completed learning for a predetermined number of epochs.
[0068] If the learning unit 193 determines that the reference value use termination condition is not satisfied (step S222: NO), the process returns to step S103. On the other hand, if it determines in step S222 that the reference value use termination condition is not satisfied (step S222: NO), the learning unit 193 determines whether a predetermined learning termination condition is satisfied (step S231). The learning termination condition here is not limited to a specific condition. For example, the learning termination condition may be a condition that the learning unit 193 has completed learning for a predetermined number of epochs. Alternatively, the learning termination condition may be a condition that the error of the obtained model is equal to or less than a predetermined error threshold.
[0069] If the learning unit 193 determines that the learning end condition is not met (step S231: NO), the process returns to step S212. On the other hand, if the learning unit 193 determines that the learning end condition is met (step S231: YES), the learning device 100 ends the process of FIG.
[0070] On the other hand, if it is determined in step S222 that the reference value use termination condition is met (step S222: YES), the learning unit 193 determines whether a predetermined learning termination condition is met (step S241). The determination made by the learning unit 193 in step S241 is the same as that in step S231.
[0071] If it is determined that the learning termination condition is not met (step S241: NO), the learning unit 193 performs learning for one epoch without setting a reference value (step S251). After step S251, the process returns to step S241. On the other hand, if the learning unit 193 determines that the learning termination condition is met in step S241 (step S241: YES), the learning device 100 ends the process of FIG. 6.
[0072] After the learning unit 193 has performed learning for a predetermined number of epochs, the reference setting unit 192 may set a reference value, and the learning unit 193 may perform learning for an additional predetermined number of epochs based on the reference value. For example, after the learning unit 193 has performed learning for 500 epochs, the reference setting unit 192 may set a reference value. Then, the learning unit 193 may perform learning for an additional 500 epochs based on the reference value.
[0073] In this case, the reference setting unit 192 may set the training loss at the epoch at which the error based on the confirmation data is smallest (i.e., the epoch at which the accuracy based on the confirmation data is largest) as the reference value. The epoch at which the error based on the confirmation data is smallest is considered to correspond to the point just before overfitting occurs, when the training loss becomes zero. By the reference setting unit 192 setting the training loss at this epoch as the reference value and the learning unit 193 performing learning based on the reference value, it is expected that overfitting can be avoided.
[0074] FIG. 7 illustrates a third example of a processing procedure for the learning device 100 to learn a model. In the processing of FIG. 7 , the learning unit 193 performs learning for a predetermined number of epochs without setting a reference value (step S301). Next, the reference setting unit 192 sets a reference value (step S302). Specifically, the reference setting unit 192 selects an epoch with the smallest index value from among the epochs already performed by the learning unit 193, and sets a reference value based on the learning result of that epoch. As described above, the index value here may be any of training loss, validation loss, test loss, training error, validation error, and test error. Furthermore, the index value used by the reference setting unit 192 to select an epoch may differ from the index value used to set the reference value. The learning unit 193 performs learning for a predetermined number of epochs based on the reference value set by the reference setting unit 192 (step S303). After step S303, the learning device 100 terminates the processing of FIG. 7 .
[0075] Alternatively, after step S303, the learning device 100 may further set a reference value and perform learning based on the reference value. For example, as in step S302, the reference setting unit 192 selects the epoch with the smallest index value from the epochs already executed by the learning unit 193, and sets a reference value based on the learning result of that epoch. As in step S303, the learning unit 193 performs learning for a predetermined number of epochs based on the reference value set by the reference setting unit 192. For example, the user may instruct the learning device 100 whether to further set a reference value and perform learning based on the reference value.
[0076] As described above, the standard setting unit 192 sets a standard value for the degree of estimation accuracy to be achieved in further machine learning, based on the evaluation function value indicating the estimation accuracy of the model generated in machine learning. The learning unit 193 learns the model so as to bring the output value of the evaluation function closer to the standard value.
[0077] The learning device 100 uses evaluation function values, making it relatively easy to set hyperparameter values for adjusting the breadth of parameter value search. Furthermore, the learning device 100 is expected to avoid overfitting by having the learning unit 193 learn a model based on reference values. Furthermore, the learning device 100 allows for the setting of reference values corresponding to examples of hyperparameter values for adjusting the breadth of parameter value search to be set to a single value at a time. In particular, the learning device 100 does not require the operation of setting multiple reference values, learning a model, and selecting one of the reference values based on the learning results. In this respect, the learning device 100 is expected to be able to relatively shorten the time required for learning without requiring computational resources such as parallel processing. Furthermore, the learning device 100 is expected to be able to set a reference value based on the evaluation function values obtained during learning by the reference setting unit 192, allowing the reference value to be set according to the model and learning situation. In this respect, it is expected that a relatively appropriate reference value can be set.
[0078] Furthermore, the reference setting unit 192 sets the reference value to a value that indicates that the degree of fit of the model to the training data is more limited than the maximum degree of fit among the possible values of the evaluation function. According to the learning device 100, the possibility of overfitting can be reduced by learning using the reference value.
[0079] Furthermore, the reference setting unit 192 sets a reference value based on an evaluation function value obtained by applying confirmation data, which is data other than data (training data) used to update the parameter values of the model, to the model. For example, in an epoch in which an evaluation function value obtained by applying data other than training data, such as a confirmation error, to the model is small, the generalization performance of the obtained model is expected to be relatively high. By setting the training error in the epoch with the smallest confirmation error as the reference value, the reference setting unit 192 sets the training error when a model with relatively high generalization performance is obtained as the reference value. By having the learning unit 193 learn the model based on this reference value, it is expected that it will be easier to search for a solution that will result in a model with relatively high generalization performance, and in this respect, it is expected that overfitting will be avoided.
[0080] Furthermore, after one epoch of model learning repeatedly performed by the learning unit 193 is completed, the reference setting unit 192 determines whether a learning result with a smaller evaluation function value than the learning result used to set the reference value has been obtained in the learning for that epoch. If it is determined that a learning result with a smaller evaluation function value has been obtained, the reference setting unit 192 updates the reference value based on the learning result for that epoch. This allows the reference setting unit 192 to update the reference value as the model learning by the learning unit 193 progresses, and is expected to enable setting an appropriate reference value according to the progress of learning.
[0081] Furthermore, the learning unit 193 continues learning the model so as to bring the evaluation function value as close to 0 as possible until a predetermined reference value use start condition is met. This allows the reference setting unit 192 to set a reference value after learning has progressed to a certain extent and the accuracy of the model has stabilized. In this regard, the learning device 100 is expected to enable the reference setting unit 192 to set an appropriate reference value.
[0082] Furthermore, after a predetermined condition for ending the use of the reference value is met, the learning unit 193 performs model learning so as to bring the evaluation function value as close as possible to 0. According to the learning device 100, learning can be performed in the final stage of model learning without setting a reference value, and in this respect, it is expected that the accuracy of the model can be made relatively high.
[0083] Furthermore, the standard setting unit 192 selects one of the predetermined number of epochs of model learning performed by the learning unit 193 based on the learning results, and sets a standard value based on the evaluation function value indicated in the learning result for the selected epoch. According to the learning device 100, the standard setting unit 192 only needs to set the standard value once, which is expected to reduce the time required for learning without requiring computational resources such as parallel processing. Furthermore, since the standard setting unit 192 sets the standard value after the model learning by the learning unit 193 has progressed to a certain extent, it is expected to be able to set an appropriate standard value.
[0084] The standard setting unit 192 also generates a restricted evaluation function. The restricted evaluation function is a function that outputs a value equal to the output value of the evaluation function in a portion of the domain of the evaluation function where the output value of the evaluation function is equal to or greater than the standard value, and outputs a value greater than the standard value in a portion where the output value of the evaluation function is smaller than the standard value. The learning unit 193 performs model learning using the restricted evaluation function set by the standard setting unit 192. In the learning device 100, the standard value can be included in the restricted evaluation function, and the learning unit 193 does not need to refer to the standard value separately from the evaluation function. In this respect, the learning device 100 is expected to impose a relatively small load on the learning unit 193.
[0085] Fig. 8 is a diagram showing an example of the configuration of an estimation device according to an embodiment. In the configuration shown in Fig. 8, the estimation device 200 includes a communication unit 210, a display unit 220, an operation input unit 230, a storage unit 280, and a control unit 290. The storage unit 280 includes a model storage unit 181. The control unit 290 includes a data acquisition unit 291 and an estimation unit 292.
[0086] The estimation device 200 performs estimation using a model trained by the learning device 100. The use of the estimation device 200 is not limited to a specific use. For example, the estimation device 200 may be configured as a face recognition device and may calculate the similarity between a face image to be authenticated and a registered face image using the model. Alternatively, the estimation device 200 may input a given sentence into the model and estimate the emotion indicated by the sentence. In this way, the estimation device 200 can be applied to various fields, such as computer vision or natural language processing.
[0087] The estimation device 200 may be configured using a computer such as a personal computer or a workstation. The estimation device 200 may be configured using the computer used as the learning device 100. Alternatively, the estimation device 200 may be configured using a computer different from the computer used as the learning device 100.
[0088] The communication unit 210 communicates with other devices. For example, the communication unit 210 may communicate with other devices to receive data of the estimation target. The display unit 220 has a display screen such as a liquid crystal panel or an LED panel, and displays various images. For example, the display unit 220 may display the estimation results obtained by the estimation device 200. The operation input unit 230 has input devices such as a keyboard and a mouse, and accepts user operations. The operation input unit 230 may accept a user operation to instruct the start of estimation.
[0089] The storage unit 280 stores various data. The storage unit 280 is configured using a storage device included in the estimation device 200. The model storage unit 181 stores a model that has been learned by the learning device 100. In this respect, the model storage unit 181 of the estimation device 200 stores the same model as the model storage unit 181 of the learning device 100. For this reason, in FIG. 8 , the reference numeral 181 is used for the model storage unit, as in FIG. 1 .
[0090] Alternatively, when the model to be learned by learning device 100 is configured as a device separate from learning device 100, the model used by estimation device 200 may also be configured as a device separate from estimation device 200. In this case, storage unit 280 may be configured without including model storage unit 181.
[0091] The control unit 290 controls each part of the estimation device 200 and performs various processes. The functions of the control unit 290 may be performed, for example, by a CPU provided in the estimation device 200 reading and executing a program from the memory unit 280.
[0092] The data acquisition unit 291 acquires data of the estimation target. For example, when the communication unit 210 receives data of the estimation target from another device, the data acquisition unit 291 extracts the data of the estimation target from the data received by the communication unit 210. The estimation unit 292 performs estimation on the estimation target acquired by the data acquisition unit 291. The estimation unit 292 inputs the data of the estimation target acquired by the data acquisition unit 291 into a model stored in the model storage unit 181, and acquires the output of the model as the estimation result.
[0093] As described above, the estimation unit 292 calculates an estimation value for the estimation target using a trained model obtained by model training using the learning device 100. The trained model obtained by model training using the learning device 100 is set based on the value of an evaluation function that indicates an evaluation of the estimation accuracy of the model, and corresponds to an example of a model updated by machine learning that is performed so as to bring the output value of the evaluation function closer to a reference value that specifies the level of estimation accuracy to be achieved by further machine learning. In the estimation device 200, it is expected that the model will not be overfitted. In this respect, it is expected that the estimation device 200 will be able to perform estimation with high accuracy.
[0094] FIG. 9 is a diagram showing another example of the configuration of a learning device according to an embodiment. In the configuration shown in FIG. 9 , the learning device 610 includes a reference setting unit 611 and a learning unit 612. In this configuration, the reference setting unit 611 sets a reference value for determining the degree of estimation accuracy to be achieved in further machine learning, based on the value of an evaluation function indicating an evaluation of the estimation accuracy of a model generated in machine learning. The learning unit 612 updates the model by performing machine learning so as to bring the output value of the evaluation function closer to the reference value. The reference setting unit 611 corresponds to an example of a reference setting means. The learning unit 612 corresponds to an example of a learning means.
[0095] In the learning device 610, the learning unit 612 is expected to avoid overfitting by training the model based on the reference value. Furthermore, the learning device 610 allows for the setting of a reference value corresponding to an example of a hyperparameter value for adjusting the breadth of parameter value search to be set to a single value at a time. In particular, the learning device 610 does not require the operation of setting multiple reference values, training the model, and selecting one of the reference values based on the training results. In this respect, the learning device 610 is expected to be able to relatively shorten the time required for training without requiring computational resources such as parallel processing. Furthermore, in the learning device 610, the reference value can be set according to the model and the learning situation by the reference setting unit 611 setting the reference value based on the evaluation function value obtained during training, and thus it is expected to be able to set a relatively appropriate reference value.
[0096] The reference setting unit 611 can be realized, for example, by using the function of the reference setting unit 192 in Fig. 1. The learning unit 612 can be realized, for example, by using the function of the learning unit 193 in Fig. 1.
[0097] 10 is a diagram showing an example of a processing procedure in a learning method according to an embodiment. The learning method shown in FIG. 10 includes setting a standard (step S611) and performing learning (step S612). In setting the standard (step S611), the computer sets a standard value for the degree of estimation accuracy of the model to be achieved in further machine learning, based on the value of an evaluation function indicating an evaluation of the estimation accuracy of the model generated in machine learning. In performing learning (step S612), the computer performs learning of the model so as to bring the output value of the evaluation function closer to the standard value.
[0098] The learning method shown in FIG. 10 is expected to avoid overfitting by training a model based on a reference value. Furthermore, the learning method shown in FIG. 10 allows for the setting of a reference value corresponding to an example of a hyperparameter value for adjusting the breadth of parameter value search to be set to a single value at a time. In particular, the learning method shown in FIG. 10 does not require the operation of setting multiple reference values, training a model, and selecting one of the reference values based on the training results. In this respect, the learning method shown in FIG. 10 is expected to be able to relatively shorten the time required for training without requiring computational resources such as parallel processing. Furthermore, the learning method shown in FIG. 10 is expected to be able to set a reference value based on an evaluation function value obtained during training, thereby allowing the reference value to be set according to the model and the learning situation. In this respect, it is expected that a relatively appropriate reference value can be set.
[0099] 11 is a schematic block diagram illustrating the configuration of a computer according to at least one embodiment. In the configuration shown in FIG. 11, a computer 700 includes a CPU 710, a main memory device 720, an auxiliary memory device 730, an interface 740, and a non-volatile recording medium 750.
[0100] One or more of the learning device 100, the estimation device 200, and the learning device 610, or a part thereof, may be implemented in the computer 700. In this case, the operation of each of the processing units described above is stored in the auxiliary storage device 730 in the form of a program. The CPU 710 reads the program from the auxiliary storage device 730, loads it into the main storage device 720, and executes the above processing in accordance with the program. The CPU 710 also allocates storage areas in the main storage device 720 corresponding to each of the above storage units in accordance with the program. Communication between each device and other devices is performed by the interface 740, which has a communication function, and performs communication under the control of the CPU 710.
[0101] When learning device 100 is implemented in computer 700, the operations of control unit 190 and each of its components are stored in the form of a program in auxiliary storage device 730. CPU 710 reads the program from auxiliary storage device 730, loads it into main storage device 720, and executes the above-described processing in accordance with the program.
[0102] Furthermore, the CPU 710 allocates storage areas corresponding to the storage unit 180 and each unit thereof in the main storage device 720 in accordance with the program. Communication by the communication unit 110 is performed by the interface 740 having a communication function and performing communication under the control of the CPU 710. Display of images by the display unit 120 is performed by the interface 740 having a display device and displaying images under the control of the CPU 710. Reception of user operations by the operation input unit 130 is performed by the interface 740 having an input device and receiving user operations.
[0103] When the estimation device 200 is implemented in a computer 700, the operations of the control unit 290 and each of its units are stored in the form of a program in an auxiliary storage device 730. The CPU 710 reads the program from the auxiliary storage device 730, loads it into the main storage device 720, and executes the above-described processing in accordance with the program.
[0104] Furthermore, the CPU 710 allocates storage areas corresponding to the storage unit 280 and each unit thereof in the main storage device 720 in accordance with the program. Communication by the communication unit 210 is performed by the interface 740 having a communication function and performing communication under the control of the CPU 710. Display of images by the display unit 220 is performed by the interface 740 having a display device and displaying images under the control of the CPU 710. Reception of user operations by the operation input unit 230 is performed by the interface 740 having an input device and receiving user operations.
[0105] When the learning device 610 is implemented in the computer 700, the operations of the reference setting unit 611 and the learning unit 612 are stored in the form of a program in the auxiliary storage device 730. The CPU 710 reads the program from the auxiliary storage device 730, loads it into the main storage device 720, and executes the above-described processing in accordance with the program.
[0106] Furthermore, CPU 710 allocates a storage area in main memory 720 for learning device 610 to perform processing in accordance with the program. Communication between learning device 610 and other devices is performed by interface 740, which has a communication function and operates under the control of CPU 710. Interaction between learning device 610 and a user is performed by interface 740, which has a display device and an input device, displaying various images under the control of CPU 710 and accepting user operations.
[0107] One or more of the above-described programs may be recorded on nonvolatile recording medium 750. In this case, interface 740 may read the programs from nonvolatile recording medium 750. Then, CPU 710 may directly execute the programs read by interface 740, or may temporarily store the programs in main storage device 720 or auxiliary storage device 730 and then execute them.
[0108] Note that a program for executing all or part of the processing performed by learning device 100, estimation device 200, and learning device 610 may be recorded on a computer-readable recording medium, and the program recorded on the recording medium may be loaded into a computer system and executed to perform the processing of each unit. Note that the term "computer system" here includes hardware such as an operating system (OS) and peripheral devices. Furthermore, the term "computer-readable recording medium" refers to portable media such as flexible disks, optical magnetic disks, read-only memories (ROMs), and compact disc read-only memories (CD-ROMs), as well as storage devices such as hard disks built into computer systems. The program may be designed to implement part of the aforementioned functions, or may be capable of implementing the aforementioned functions in combination with a program already recorded on the computer system.
[0109] Although an embodiment of the present invention has been described above in detail with reference to the drawings, the specific configuration is not limited to this embodiment, and includes designs within the scope of the gist of the present invention.
[0110] Some or all of the above embodiments can be described as, but are not limited to, the following supplementary notes.
[0111] (Supplementary Note 1) A learning device comprising: a standard setting means for setting a standard value for the degree of estimation accuracy to be achieved in further machine learning based on the value of an evaluation function that indicates an evaluation of the estimation accuracy of a model generated in machine learning; and a learning means for performing machine learning to update the model so as to bring the output value of the evaluation function closer to the standard value.
[0112] (Supplementary Note 2) The learning device according to Supplementary Note 1, wherein the standard setting means sets the standard value to a value that indicates that the fitness of the model to training data, which is data for updating parameter values of the model, is more limited than the maximum fitness, among the values that the evaluation function can take.
[0113] (Supplementary Note 3) The learning device according to Supplementary Note 1 or Supplementary Note 2, wherein the standard setting means sets the standard value based on an evaluation function value obtained by applying confirmation data, which is data other than data for updating parameter values of the model, to the model.
[0114] (Supplementary Note 4) The learning device described in any one of Supplementary Notes 1 to 3, wherein after one unit of learning of the model that is repeatedly performed by the learning means is completed, the standard setting means determines whether or not a learning result has been obtained in that one unit of learning that indicates that the evaluation function value shows a higher degree of fit of the model to the input data than the learning result used to set the standard value, and if it is determined that a learning result showing a higher degree of fit has been obtained, updates the standard value based on the result of that one unit of learning.
[0115] (Supplementary Note 5) The learning device according to Supplementary Note 4, wherein the learning means learns the model so that the output value of the evaluation function approaches the evaluation function value that indicates the maximum degree of fitness until a predetermined reference value use start condition is met.
[0116] (Supplementary Note 6) The learning device according to Supplementary Note 4 or Supplementary Note 5, wherein the learning means, after a predetermined condition for ending use of the reference value is met, learns the model so as to bring the output value of the evaluation function closer to the evaluation function value indicating the maximum degree of fitness.
[0117] (Supplementary Note 7) The learning device described in any one of Supplementary Notes 1 to 3, wherein the standard setting means selects any epoch from the predetermined number of epochs of learning of the model performed by the learning means based on the learning results, and sets the standard value based on the evaluation function value indicated in the learning result at the selected epoch.
[0118] (Supplementary Note 8) The learning device described in any one of Supplementary Notes 1 to 7, wherein the standard setting means generates a restricted evaluation function that outputs a value equal to the output value of the evaluation function in a portion of the domain of the evaluation function where the output value of the evaluation function is equal to the standard value or where the output value indicates that the fitness of the model to training data, which is data for updating parameter values of the model, is smaller than the standard value, and outputs a value indicating that the fitness is smaller than the standard value in a portion where the output value of the evaluation function indicates that the fitness is larger than the standard value; and the learning means learns the model using the restricted evaluation function.
[0119] (Supplementary Note 9) An estimation device comprising: an estimation unit that calculates an estimated value for an estimation target using the model that has been updated through machine learning, the estimate being set based on a value of an evaluation function that indicates an evaluation of the estimation accuracy of a model generated through machine learning, and that is performed so as to bring an output value of the evaluation function closer to a reference value that specifies how much estimation accuracy the model should achieve through further machine learning.
[0120] (Supplementary Note 10) A learning method including: a computer sets a reference value for the degree of estimation accuracy to be achieved in further machine learning based on a value of an evaluation function indicating an evaluation of the estimation accuracy of a model generated in machine learning; and learns the model so that the output value of the evaluation function approaches the reference value.
[0121] (Supplementary Note 11) A recording medium for recording a program for causing a computer to execute the following steps: setting a reference value for the degree of estimation accuracy to be achieved in further machine learning, based on the value of an evaluation function indicating an evaluation of the estimation accuracy of a model generated in machine learning; and training the model so as to bring the output value of the evaluation function closer to the reference value.
[0122] The present invention may be applied to a learning device, an estimation device, a learning method, and a recording medium.
[0123] 100, 610 Learning device 110, 210 Communication unit 120, 220 Display unit 130, 230 Operation input unit 180, 280 Storage unit 181 Model storage unit 190, 290 Control unit 191, 291 Data acquisition unit 192, 611 Reference setting unit 193, 612 Learning unit 292 Estimation unit
Claims
1. A standard setting means for setting a standard value for the degree of estimation accuracy of the model to be obtained in further machine learning based on a value of an evaluation function indicating an evaluation of the estimation accuracy of the model generated in machine learning; a learning means for performing machine learning to update the model so as to bring the output value of the evaluation function closer to the reference value; A learning device comprising:
2. the standard setting means sets the standard value to a value indicating that the degree of conformance of the model to training data, which is data for updating parameter values of the model, is limited to less than the maximum degree of conformance, among values that the evaluation function can take. The learning device according to claim 1 .
3. the standard setting means sets the standard value based on an evaluation function value obtained by applying confirmation data, which is data other than data for updating parameter values of the model, to the model; The learning device according to claim 1 or 2.
4. the standard setting means, after completing one unit of learning of the model which is repeatedly performed by the learning means, determines whether or not a learning result indicating that the model has a higher degree of fitness for the input data in terms of an evaluation function value has been obtained in the one unit of learning, compared to the learning result used for setting the standard value, and updates the standard value based on the result of the one unit of learning if it is determined that a learning result indicating a higher degree of fitness has been obtained. A learning device according to any one of claims 1 to 3.
5. the learning means performs learning of the model so as to bring the output value of the evaluation function closer to an evaluation function value showing a maximum degree of compatibility until a predetermined reference value use start condition is satisfied. The learning device according to claim 4.
6. the learning means, after a predetermined reference value use termination condition is satisfied, performs learning of the model so as to bring the output value of the evaluation function closer to an evaluation function value showing a maximum degree of compatibility. The learning device according to claim 4 or 5.
7. the standard setting means selects one of the epochs of the model learning performed by the learning means based on a learning result, and sets the standard value based on an evaluation function value indicated in the learning result in the selected epoch. A learning device according to any one of claims 1 to 3.
8. An estimation unit that calculates an estimation value for an estimation target using the model updated by machine learning that is performed so as to bring an output value of the evaluation function closer to a reference value that is set based on an evaluation of the estimation accuracy of a model generated in machine learning and that specifies the degree of estimation accuracy of the model to be obtained in further machine learning. An estimation device comprising:
9. The computer Based on a value of an evaluation function indicating an evaluation of the estimation accuracy of the model generated in the machine learning, a reference value is set as to how much estimation accuracy of the model is to be obtained in further machine learning; The model is trained so that the output value of the evaluation function approaches the reference value. A learning method that includes:
10. On the computer, Based on a value of an evaluation function indicating an evaluation of the estimation accuracy of the model generated in the machine learning, a reference value is set for the degree of estimation accuracy of the model to be obtained in further machine learning; learning the model so as to bring the output value of the evaluation function closer to the reference value; A program for executing.
