Learning device, learning method, learning program, and learning data compression device

Abstract

To provide a learning device, a method and a program that can autonomously add a prototype.SOLUTION: In a learning device, a processing unit comprises: a learning data acquisition section that acquires learning data which includes input feature amount data indicating input feature amounts and teacher label data indicating teacher labels associated with the input feature amounts; a predicted label acquisition section that acquires predicted labels output by a learning model in which the input feature amounts are input; a prototype feature amount update section that updates prototype feature amounts on the basis of a coincidence degree between at least each teacher label and each predicted label; an update frequency increase section that increments an update frequency corresponding to update frequency information of a prototype in which a prototype feature amount is updated, by one; and a prototype addition section that determines whether or not to add the prototype to a storage section on the basis of a certainty degree indicating accuracy of determination by the leaning model to output the teacher label associated with the input feature amount as the predicted label according to input of the input feature amount, the teacher label and the predicted label.SELECTED DRAWING: Figure 2

Description

[Technical Field]

[0001] The present invention relates to a learning device, a learning method, a learning program, and a learning data compression device. [Background technology]

[0002] It has been known for some time that information is encoded using Learning Vector Quantization (LVQ) in image processing, audio processing, etc. LVQ is a supervised learning model that learns a prototype vector that assigns a label to an input vector, and as shown in Patent Document 1, for example, it can be used as a classifier that outputs a label for an input vector, and is therefore utilized in pattern recognition, etc.

[0003] In machine learning using LVQ, an LVQ neural network is used that has an input layer, a competitive layer, and an output layer. The input layer stores input vectors. The competitive layer has prototype vectors, which are feature vectors that represent each label, and labels, and outputs the input vector obtained from the input layer and the nearest prototype vector. The output layer linearly outputs the labels of the prototype vectors obtained from the competitive layer.

[0004] As shown in Non-Patent Document 1, the learning algorithm using an LVQ neural network is executed, for example, in the following steps (1) to (4). (1) An input vector corresponding to the input feature of the training data is input to the input layer. (2) In the competitive layer, the label of the prototype vector nearest to the input vector obtained from the input layer is determined as the “winner.” (3) If the label of the input vector and the label of the prototype vector determined as the "winner" are the same, the prototype vector is moved closer to the input vector according to the learning rate. If the label of the input vector x and the label of the prototype vector determined as the "winner" are different, the prototype vector is moved farther away from the input vector according to the learning rate. (4) Repeat the above steps (1) to (3) to learn the prototype vector. [Prior art documents] [Patent documents]

[0005] [Patent Document 1] Japanese Patent Application Laid-Open No. 2004-110196 [Non-patent literature]

[0006] [Non-Patent Document 1] T. Kohonen, “Improved versions of learning vector quantization,” in Proceedings of the IEEE International Joint Conference on Neural Networks, pp.545-550, 1990 Summary of the Invention [Problem to be solved by the invention]

[0007] In learning models that enable online machine learning using a learning rate, such as LVQ, prototypes stored in a learning device are sequentially added by an operator according to the learning accuracy. However, if the number of prototypes stored in the learning device is small, the learning accuracy will be low. On the other hand, if the number of prototypes stored in the learning device is large, or if many prototypes are added to the learning device, the learning device may become unable to learn if the storage capacity of the learning device is exceeded. Furthermore, if the number of prototypes stored in the learning device is large, or if many prototypes are added to the learning device, the time required for the learning process by the learning device may be long, and the learning efficiency may be reduced.

[0008] The present invention has been made to solve these problems, and aims to provide a learning device, a learning method, and a learning program that enable the construction of a machine learning model with high predictive accuracy using fewer learning resources and less learning time by deleting prototypes. [Means for solving the problem]

[0009] The learning device of the present invention includes a memory unit that stores multiple prototypes, each of which includes a prototype label, a prototype feature corresponding to the prototype label, and update count information indicating the number of times the prototype feature has been updated; a learning data acquisition unit that acquires learning data including input feature data indicating the input features and teacher label data indicating the teacher label associated with the input features; a predicted label acquisition unit that acquires a predicted label output by a learning model to which the input features have been input; a prototype feature update unit that updates the prototype features based on at least the degree of agreement between the teacher label and the predicted label; an update count increment unit that increments by one the update count corresponding to the update count information of the prototype whose prototype features have been updated; and a prototype addition unit that decides whether to add a prototype to the memory unit based on a confidence factor indicating the likelihood that the learning model will output the teacher label associated with the input features as a predicted label in response to the input of the input features, the teacher label, and the predicted label.

[0010] Furthermore, it is preferable that the learning device of the present invention further includes a prototype extraction unit that extracts prototypes whose update counts satisfy predetermined conditions, and a prototype information output unit that outputs prototype information indicating the prototypes extracted by the prototype extraction unit.

[0011] Furthermore, in the learning device according to the present invention, it is preferable that the prototype extraction unit extracts a prototype whose number of updates is less than a predetermined number threshold.

[0012] Furthermore, in the learning device according to the present invention, it is preferable that the prototype extraction unit extracts a predetermined number of prototypes in ascending order of the number of updates.

[0013] Furthermore, it is preferable that the learning device of the present invention further includes an image generation unit that generates prototype image data showing a prototype image that displays each prototype in a manner corresponding to the number of updates, and a prototype information output unit that outputs the prototype image data generated by the image generation unit.

[0014] Furthermore, it is preferable that the learning device according to the present invention further comprises a prototype deletion unit that deletes the prototype that has been updated the least number of times, and a deletion determination unit that determines whether or not to terminate the deletion of prototypes by the prototype deletion unit based on the confidence level.

[0015] In addition, the learning method of the present invention is a learning method for training a learning model in a learning device having a memory unit that stores multiple prototypes, each of which includes a prototype label, a prototype feature corresponding to the prototype label, and update count information indicating the number of times the prototype feature has been updated, and includes the steps of: acquiring learning data including input feature data indicating the input features and teacher label data indicating the teacher label associated with the input features; acquiring a predicted label output by a learning model to which the input features have been input; updating the prototype features based on at least the degree of match between the teacher label and the predicted label; incrementing the number of updates corresponding to the update count information of the prototype whose prototype features have been updated by 1; and determining whether to add the prototype to the memory unit based on a confidence level indicating the likelihood that the learning model will output the teacher label associated with the input features as a predicted label in response to the input of the input features, the teacher label, and the predicted label.

[0016] The control program of the present invention is a control program for causing a learning device having a memory unit that stores multiple prototypes, each of which includes a prototype label, a prototype feature corresponding to the prototype label, and update count information indicating the number of times the prototype feature has been updated, to learn a learning model.The control program causes the learning device to execute the following processes: acquire learning data including input feature data indicating the input features and teacher label data indicating the teacher label associated with the input features; acquire a predicted label output by a learning model to which the input features have been input; update the prototype features based on at least the degree of agreement between the teacher label and the predicted label; increment the number of updates corresponding to the update count information of the prototype whose prototype features have been updated by 1; and determine whether to add the prototype to the memory unit based on the teacher label and the predicted label, the confidence level indicating the probability that the learning model will output the teacher label associated with the input features as a predicted label in response to the input of the input features.

[0017] The training data compression device of the present invention includes a memory unit that stores a prototype label and a plurality of prototypes, each of which includes a prototype feature corresponding to the prototype label; a training data acquisition unit that acquires training data including input feature data indicating the input feature and teacher label data indicating the teacher label associated with the input feature; a predicted label acquisition unit that acquires a predicted label output by a training model to which the input feature is input; and a training data addition unit that decides whether to add the training data to the memory unit based on a confidence factor indicating the likelihood that the learning model will output the teacher label associated with the input feature as a predicted label in response to the input of the input feature, the teacher label, and the predicted label. [Effects of the Invention]

[0018] The learning device, learning method, and learning program according to the present invention can build a machine learning model with high prediction accuracy using fewer learning resources and less learning time. [Brief explanation of the drawings]

[0019] [Figure 1]1A and 1B are explanatory diagrams for explaining an overview of a learning device according to an embodiment, in which (a) is an example of an image including generated prototype information, (b) is another example of an image including generated prototype information, and (c) is a diagram showing an example of a prototype image after the prototype has been deleted by the learning device. [Figure 2] FIG. 1 is a block diagram of a learning device according to a first embodiment. [Figure 3] FIG. 3 is a schematic diagram of the learning model shown in FIG. 2. [Figure 4] FIG. 3 is a diagram showing a prototype table shown in FIG. 2. [Figure 5] FIG. 3 is a diagram showing a certainty factor table shown in FIG. [Figure 6] FIG. 3 is a diagram illustrating a learning rate table shown in FIG. [Figure 7] 3 is a flowchart showing a learning process performed by the learning device shown in FIG. 2. [Figure 8] 8 is a flowchart showing more detailed processing of the processing of S106 shown in FIG. 7. [Figure 9] 3 is a flowchart showing a prototype deletion process performed by the learning device shown in FIG. 2. [Figure 10] (a) is a first example of prototype information displayed on the output unit shown in Figure 2, (b) is a second example of prototype information displayed on the output unit shown in Figure 2, (c) is a third example of prototype information displayed on the output unit shown in Figure 2, and (d) is a fourth example of prototype information displayed on the output unit shown in Figure 2. [Figure 11] 2. (a) is a fifth example of prototype information displayed in the output section shown in FIG. 2, and (b) is a sixth example of prototype information displayed in the output section shown in FIG. [Figure 12] FIG. 10 is a block diagram of a learning device according to a second embodiment. [Figure 13] 13 is a flowchart showing a prototype deletion process performed by the learning device shown in FIG. 12. [Figure 14](a) is a first example of a prototype image displayed on the output unit shown in Figure 12, (b) is a second example of a prototype image displayed on the output unit shown in Figure 12, (c) is a third example of a prototype image displayed on the output unit shown in Figure 12, and (d) is a fourth example of a prototype image displayed on the output unit shown in Figure 12. [Figure 15] FIG. 10 is a block diagram of a learning device according to a third embodiment. [Figure 16] 16 is a flowchart showing a prototype deletion process performed by the learning device shown in FIG. 15. DETAILED DESCRIPTION OF THE INVENTION

[0020] Various embodiments of the present invention will be described below with reference to the drawings. However, it should be noted that the technical scope of the present invention is not limited to these embodiments, but extends to the inventions set forth in the claims and their equivalents.

[0021] (Overview of the learning device according to the embodiment) 1 is an explanatory diagram for explaining an overview of a learning device according to an embodiment. The learning device according to an embodiment is, for example, a personal computer (PC). The learning device according to an embodiment may also be an electronic computer such as a server device, a tablet terminal, or a tablet PC.

[0022] The learning device performs a learning process to train a learning model while updating prototype features stored in a storage unit using learning data including input feature data indicating input features and teacher label data indicating teacher labels associated with the input features. The learning device increments by 1 the update count corresponding to the prototype update count information for the updated prototype features. The learning device determines whether to add a new prototype to the storage unit based on the teacher label, the confidence level indicating the likelihood that the learning model will output the teacher label associated with the input features as a predicted label in response to the input of the input features, and the predicted label. More specifically, the learning device determines to add the learning data to the storage unit as a new prototype when the confidence level of the teacher label is greater than a predetermined threshold and the predicted label does not match the teacher label. The learning device can autonomously add prototypes by determining whether to add a new prototype based on the confidence level and the predicted label. The learning device performs various processes to delete prototypes based on the update count information indicating the number of times each prototype has been updated.

[0023] In one example, the learning device extracts prototypes whose update counts satisfy a predetermined condition and outputs prototype information indicating the extracted prototypes. The learning device may extract prototypes whose update counts are less than a predetermined threshold, or may extract a predetermined number of prototypes in ascending order of update count. The learning device outputs the prototype information indicating the extracted prototypes to a display device, and the display device, to which the prototype information has been input, displays an image including the prototype information.

[0024] FIG. 1(a) is a diagram showing an example of an image including prototype information.

[0025] Prototype image 101 displays 30 pieces of prototype information with the least number of updates for each of the 10 numbers from "0" to "9" in the MNIST database. The numbers displayed in prototype image 101 differ from the fontified numbers, and for example, "1" and "7" contain images of prototype information that are similar to each other. The operator selects prototype information from the prototype information included in prototype image 101 that differs from the shape of the fontified numbers and deletes the selected prototype information.

[0026] In another example, the learning device generates prototype image data representing a prototype image that displays each of the prototypes in a manner corresponding to the number of updates, and outputs the generated prototype image data. The learning device outputs the generated prototype image data to a display device, and the display device to which the prototype image data has been input displays the prototype image that displays each of the prototypes in a manner corresponding to the number of updates.

[0027] Fig. 1(b) shows another example of a generated prototype image. In Fig. 1(b), prototype image 102a is an image containing training data used by the learning device, and prototype image 102b is a prototype image containing a prototype generated by the learning device.

[0028] The learning device displays each prototype in a size that corresponds to the number of times it has been updated. Prototypes that have been updated many times are displayed larger, and prototypes that have been updated fewer times are displayed smaller. The operator selects prototype information displayed smaller from the prototype information included in prototype image 102b and deletes the selected prototype information.

[0029] In yet another example, the learning device deletes the prototype with the least number of updates and determines whether to terminate the prototype deletion by the prototype deletion unit based on the confidence factor, where the learning device determines to terminate the prototype deletion by the prototype deletion unit when the confidence factor after the prototype deletion is higher than the confidence factor before the prototype deletion.

[0030] Fig. 1(c) shows an example of a prototype image after the prototype has been deleted by the learning device. In Fig. 1(c), prototype image 103a is an image containing training data used by the learning device, and prototype image 103b is a prototype image showing the prototype generated by the learning device.

[0031] The learning device according to the embodiment can remove prototypes that have been updated infrequently, thereby removing prototypes to which erroneous label data has been assigned, thereby reducing the likelihood of the machine learning model making erroneous predictions. Furthermore, the learning device according to the embodiment can dynamically remove unnecessary training data from the training data, thereby improving the quality of the training data and enabling the construction of a machine learning model with high prediction accuracy.

[0032] Furthermore, the learning device according to the embodiment may be used as a training data compression device that compresses the training data to be used. Conventional learning devices allow an operator to visually check the training data and remove unnecessary training data. However, when adding training data, the operator must sequentially check the training data and remove training data that the operator determines to be unnecessary. In contrast, the learning device according to the embodiment can dynamically remove unnecessary training data by adding the training data to the storage unit 12 as training data to be included in the compressed training data based on the confidence factor, teacher label, and predicted label. Furthermore, when adding training data, the learning device according to the embodiment checks the update count of the prototype most similar to the training data to be added, thereby automatically outputting low-confidence training data and reducing the cost of removing unnecessary training data.

[0033] (Configuration and Functions of the Learning Device According to the First Embodiment) FIG. 2 is a block diagram of the learning device according to the first embodiment.

[0034] The learning device 1 includes a communication unit 11, a storage unit 12, an input unit 13, an output unit 14, and a processing unit 20. The learning device 1 executes a learning process to train a learning model by updating prototype features stored in the storage unit using learning data including input feature data indicating input features and teacher label data indicating teacher labels associated with the input features. The communication unit 11, the storage unit 12, the input unit 13, the output unit 14, and the processing unit 20 are communicatively connected via a bus 15.

[0035] The communication unit 11 performs short-range wireless communication according to the Bluetooth (registered trademark) communication method, establishes terminal-to-terminal wireless communication with an external device, and performs direct wireless communication. The communication method for short-range wireless communication is not limited to Bluetooth (registered trademark), and may be, for example, Bluetooth (registered trademark) Low Energy. Furthermore, the communication unit 11 is not limited to one for performing short-range wireless communication, and may be one that receives various signals via, for example, infrared communication. In this case, the communication unit 11 may generate data by modulating a predetermined radio wave received from the external device into a specific signal. Furthermore, the communication unit 11 may be a communication interface including an interface for connecting to a USB (Universal Serial Bus) or the like, a communication interface for a wired or wireless LAN (Local Area Network), or the like.

[0036] The storage unit 12 includes at least one of a semiconductor memory device such as a read-only memory (ROM) or a random access memory (RAM), a magnetic tape device, a magnetic disk device, or an optical disk device. The storage unit 12 stores an operating system program, a driver program, a control program, data, and the like used for processing by the processing unit 20. The driver programs stored in the storage unit 12 include a communication device driver program that controls the communication unit 11, an input device driver program that controls the input unit 13, and an output device driver program that controls the output unit 14. The control programs stored in the storage unit 12 include application programs for executing various processes described below. The various programs stored in the storage unit 12 may be installed in the storage unit 12 from a computer-readable portable recording medium such as a CD-ROM or a DVD-ROM using a known setup program or the like.

[0037] The storage unit 12 stores, for example, a learning model 16, a prototype table 17 indicating a plurality of prototypes, a confidence level table 18, and a learning rate table 19. The storage unit 12 may also temporarily store data related to a predetermined process.

[0038] FIG. 3 is a schematic diagram of the learning model 16.

[0039] The learning model 16 has an input layer, a competitive layer, and an output layer, and outputs a predicted label in response to an input vector x, which is an example of an input feature. At least one artificial neuron (hereinafter simply referred to as a "neuron") is set in the input layer, competitive layer, and output layer.

[0040] The input layer receives an input vector x corresponding to the input feature data included in the training data. The input feature data is a data group containing multiple pieces of data arranged in a predetermined data sequence. When the input feature data is image data with a total of n pixels, the input vector x{x1, x2,...x n} indicates the color value of each pixel in the image data, x1, x2,...x n The input vector x can be any vector. For example, if the input feature data are four types of attribute parameter values ​​related to a flower, the input vector x{x1, x2, x3, x4} is the petal length value x1, the petal width value x2, the sepal length value x3, and the sepal width value x4.

[0041] The competitive layer includes neurons corresponding to the prototype vectors. The number of neurons included in the competitive layer is equal to or greater than the number (k) of preset labels. For example, the prototype vector of the ith label among the labels is m i {μ1,μ2,μ n}. In each of the k neurons in the competitive layer, the input vector x{x1, x2, . . . x n} and the prototype vector m i {μ1,μ2,μ n} is calculated. i The number of dimensions of the prototype vector m is the same as the number of dimensions of the input vector x. i μ1, μ2, μ in n The initial value of can be any value.

[0042] The output layer defines the prototype label corresponding to the shortest distance among the distances output from each neuron in the competitive layer as the predicted label L(m j )

[0043] FIG. 4 is a diagram showing the prototype table 17.

[0044] The prototype table 17 contains the prototype label L(m i ), prototype vector m i , update count information n i , and co-occurrence frequency information (a i , b i , c i and d i) is stored in association with an identifier for each prototype. i is the prototype feature of the prototype obtained by the learning process by the learning device 1, that is, the prototype vector m i The prototype associated with identifier "1" corresponds to "apple," and the prototype associated with identifier "2" corresponds to "orange." The prototype associated with identifier "3" corresponds to "grapes," and the prototype associated with identifier "4" corresponds to "none." While prototype table 17 stores four prototypes, the number of prototypes stored in the prototype table may be any number. The manner in which prototypes are stored in storage unit 12 is not limited to prototype table 17.

[0045] FIG. 5 is a diagram showing the confidence level table 18. As shown in FIG.

[0046] Confidence Table 18 shows the confidence R i are stored for each prototype in association with the same identifier as the identifier in the prototype table 17. The confidence R1 associated with the identifier "1" is the learning rate corresponding to "apple", and the confidence R2 associated with the identifier "2" is the learning rate corresponding to "orange". Furthermore, the confidence R3 associated with the identifier "3" is the learning rate corresponding to "grapes", and the confidence R4 associated with the identifier "4" is the learning rate corresponding to "none". The confidence table 18 stores the four confidence R i The confidence R stored in the confidence table 18 is i The number of prototypes may be the same as the number of prototypes stored in the prototype table 17.

[0047] FIG. 6 is a diagram showing the learning rate table 19. As shown in FIG.

[0048] Learning rate table 19 shows the learning rate α iare stored for each prototype in association with the same identifier as the identifier in the prototype table 17 and the confidence table 18. The learning rate α1 associated with the identifier "1" is the learning rate corresponding to "apples", and the learning rate α2 associated with the identifier "2" is the learning rate corresponding to "oranges". Furthermore, the learning rate α3 associated with the identifier "3" is the learning rate corresponding to "grapes", and the learning rate α4 associated with the identifier "4" is the learning rate corresponding to "none". The learning rate table 19 stores four learning rates α i However, the learning rate α i The number of prototypes may be the same as the number of prototypes stored in the prototype table 17.

[0049] The input unit 13 is, for example, a keyboard, a mouse, or a pointing device such as a touch panel. A user can use the input unit 13 to input letters, numbers, and symbols, or positions on the display screen of the output unit 14, etc. When operated by a user, the input unit 13 generates a signal corresponding to the operation. The generated signal is then supplied to the processing unit 20 as an instruction from the user.

[0050] The output unit 14 is a liquid crystal display. The output unit 14 may also be an organic EL (Electro-Luminescence) display, etc. The output unit 14 displays a video corresponding to the video data supplied from the processing unit 20, an image corresponding to the image data, etc.

[0051] The processing unit 20 includes one or more processors and their peripheral circuits. The processing unit 20 controls the overall operation of the learning device 1 and is, for example, a CPU (Central Processing Unit). The processing unit 20 executes various information processes in an appropriate order based on the programs stored in the storage unit 12 and various instructions input by the user via the input unit 13, and also controls the operation of the communication unit 11 and the output unit 14. The processing unit 20 executes various information processes based on the operating system program, driver program, and control program stored in the storage unit 12. The processing unit 20 can also execute multiple programs in parallel.

[0052] The processing unit 20 includes a learning data acquisition unit 21, a predicted label acquisition unit 22, a prototype addition unit 23, a co-occurrence frequency change unit 24, a confidence factor calculation unit 25, a learning rate calculation unit 26, a prototype feature update unit 27, and an update count increase unit 28. The processing unit 20 further includes a prototype extraction unit 31, a prototype information output unit 32, and a prototype deletion unit 33. Each of these units is a functional module realized by a program executed by a processor included in the processing unit 20. Alternatively, each of these units may be implemented in the learning device 1 as firmware.

[0053] (Learning process by learning device 1) Fig. 7 is a flowchart showing the learning process by the learning device 1. The learning process shown in Fig. 7 is executed mainly by the processing unit 20 in cooperation with each element of the learning device 1, based on a control program stored in advance in the storage unit 12.

[0054] First, the training data acquiring unit 21 acquires training data including input feature data indicating input feature amounts and teacher label data indicating teacher labels associated with the input feature amounts (S101). Specifically, the training data acquiring unit 21 acquires training data including input feature data indicating an input vector x and teacher label data indicating a teacher label L(x) associated with the input vector x.

[0055] Next, the predicted label acquisition unit 22 uses the learning model 16 to obtain the predicted label L(m j The predicted label acquisition unit 22 inputs the input vector x to the input layer of the learning model 16, and obtains the predicted label L(m j ) is obtained.

[0056] Next, the prototype adding unit 23 calculates the predicted label L(m j ) matches the teacher label L(x) obtained in S101 (S103).

[0057] The prototype adding unit 23 adds the predicted label L(m j ) does not match the teacher label L(x) (S103-NO), the confidence level R i is a given threshold R TH The prototype adding unit 23 determines whether the confidence factor R i , teacher label L(m j ), and the predicted label L(x), it is determined whether to add the prototype to the storage unit 12.

[0058] The prototype adding unit 23 adds the confidence R i is a given threshold R TH If it is determined that the difference is greater than the threshold (S104-YES), a new prototype is added (S105). The prototype vector of the prototype to be added is the input vector x corresponding to the input feature data acquired in the process of S101. The prototype label of the prototype to be added is the teacher label L(x) corresponding to the teacher label data acquired in the process of S101. When the process of S105 ends, the learning process by the learning device 1 ends without executing the processes of S106 to S110.

[0059] The predicted label L(m j ) is determined to match the teacher label L(x) (S103-YES), or the confidence R iis a given threshold R TH If it is determined that the number is equal to or less than the predetermined number (S104-NO), the process proceeds to S106.

[0060] Next, the co-occurrence frequency change unit 24 calculates the teacher label L(x) and the predicted label L(m j ) and prototype label L(m i The co-occurrence frequency change unit 24 changes the co-occurrence frequency information based on the degree of coincidence between the teacher label L(x) and the predicted label L(m j ) and prototype label L(m i The co-occurrence frequency information is changed so that the information indicating whether or not the two words match is reflected in the co-occurrence frequency information.

[0061] FIG. 8 is a flowchart showing the process of S106 shown in FIG. 7 in more detail.

[0062] First, the co-occurrence frequency change unit 24 calculates the prototype label L(m i ) from the prototype table 17 (S201). Next, the co-occurrence frequency change unit 24 calculates the prototype label L(m i ) and the predicted label L(m j ) match (S202).

[0063] The co-occurrence frequency change unit 24 changes the prototype label L(m i ) and predicted label L(m j ) match (S202-YES), the predicted label L(m j ) and the teacher label L(x) acquired in S101 (S203).

[0064] The co-occurrence frequency change unit 24 calculates the predicted label L(m j ) and the teacher label L(x) match (S203-YES), and the first co-occurrence frequency a i The co-occurrence frequency change unit 24 adds a value "1" to the predicted label L(m j) and the teacher label L(x) do not match (S203-NO), and the second co-occurrence frequency b i The value "1" is added to (S205).

[0065] The co-occurrence frequency change unit 24 changes the prototype label L(m i ) and predicted label L(m j ) does not match (S202-NO), the predicted label L(m j ) and the teacher label L(x) acquired in S101 (S206).

[0066] The co-occurrence frequency change unit 24 calculates the predicted label L(m j ) and the teacher label L(x) are determined to match (S206-YES), and the third co-occurrence frequency c i The co-occurrence frequency change unit 24 adds a value "1" to the predicted label L(m j ) and the teacher label L(x) do not match (S206-NO), and the fourth co-occurrence frequency d i The value "1" is added to (S208).

[0067] When the process of S106 is completed, the confidence factor calculation unit 25 calculates the confidence factor R based on the co-occurrence frequency information changed in S106. i The confidence calculation unit 25 calculates the confidence R using a causal inference model, which is an equation that indicates the causal relationship between two events (S107). i Specifically, the confidence calculation unit 25 calculates the confidence R for each co-occurrence frequency stored in the prototype table 17 using the following formula (1): i Calculate the following.

[0068]

number

[0069] Next, the learning rate calculation unit 26 calculates the confidence factor R calculated in the process of S107. i From learning rate α i (S108), and the calculated learning rate α iis stored in the learning rate table 19. Specifically, the learning rate calculation unit 26 calculates a learning rate α i Calculate the following.

[0070]

number

[0071] When the process of S107 is completed, the learning rate calculation unit 26 calculates all the learning rates α i The learning rate calculation unit 26 determines whether all the learning rates α i The processes of S106 to S109 are repeated until it is determined that the learning rates α1 to α4 corresponding to "apples," "tangerines," "grapes," and "pears" have been updated (S109-YES). By repeating the processes of S106 to S109, the learning rates α1 to α4 corresponding to "apples," "tangerines," "grapes," and "pears," respectively, are updated sequentially.

[0072] All learning rates α in learning rate table 19 i has been updated (S109-YES), the prototype feature update unit 27 updates the prototype features stored in the prototype table 17 based on the learning rate updated in the process of S108 (S110). Specifically, the prototype feature update unit 27 updates the prototype vector m i The learning rate α updated in the process of S107 i Based on this, it is updated using equation (3).

[0073]

number

[0074] Next, the update count increment unit 28 increments the update count corresponding to the update count information of the prototype whose prototype feature has been updated by 1 (S101). In the prototype table 17, the update count corresponding to the update count information of the prototype whose prototype feature has been updated is incremented by 1.

[0075] The update count increment unit 28 increments the update count corresponding to the update count information of the prototype whose prototype feature has been updated by 1, thereby completing the learning process of the learning model 16 using the learning data.

[0076] (Prototype deletion process by learning device 1) Figure 9 is a flowchart showing the prototype deletion process by the learning device 1. The learning process shown in Figure 9 is executed mainly by processing unit 20 in cooperation with each element of learning device 1, based on a control program stored in advance in storage unit 12.

[0077] First, the prototype extraction unit 31 acquires a prototype deletion instruction (S301). The prototype deletion instruction is input via the input unit 13 by, for example, an operator operating the learning device 1.

[0078] Next, the prototype extraction unit 31 acquires extraction condition information (S302). The prototype extraction conditions are input via the input unit 13 by the operator operating the learning device 1, similar to the prototype deletion instruction. The prototype extraction conditions are defined based on the number of updates corresponding to the update count information. The prototype extraction condition may be that the number of updates is less than a predetermined threshold value, or that the number of updates is greater than a predetermined threshold value. Furthermore, the prototype extraction condition may be a predetermined number in descending order of the number of updates, or a predetermined number in descending order of the number of updates.

[0079] Next, the prototype extraction unit 31 extracts prototypes whose update counts satisfy the prototype extraction conditions acquired in the process of S302 (S303). The prototype extraction unit 31 references the prototype table 17 and extracts prototypes whose update counts corresponding to the update count information satisfy the prototype extraction conditions acquired in the process of S302. The prototype extraction unit 31 stores the extracted prototypes in the storage unit 12.

[0080] Next, the prototype information output unit 32 outputs prototype information indicating the prototypes extracted by the prototype extraction unit 31 to the output unit 14 (S304). The prototype information is, for example, image data that displays selectable images of the prototypes extracted by the prototype extraction unit 31. The output unit 14 displays the input prototype information.

[0081] Figure 10(a) is a first example of prototype information displayed on output unit 14, and Figure 10(b) is a second example of prototype information displayed on output unit 14. Figure 10(c) is a third example of prototype information displayed on output unit 14, and Figure 10(d) is a fourth example of prototype information displayed on output unit 14. Figure 11(a) is a fifth example of prototype information displayed on output unit 14, and Figure 11(b) is a sixth example of prototype information displayed on output unit 14.

[0082] Prototype information 201 and 202 are examples of learning using images of the numbers "0" through "9" from the MNIST database as training data. Prototype information 203 and 204 are examples of learning using images contained in a database of fashion standards, also known as "Fassion-MNIST," as training data, in which the prototypes are clothing such as T-shirts, pants, bags, and shoes. Prototype information 205 and 206 are examples of learning using images of the letters "a" through "z" contained in a database of handwritten alphabets, also known as "MNIST Letters," as training data. Prototype information 201, 203, and 205 display 30 pieces of prototype information in descending order of the number of updates, while prototype information 202, 204, and 206 display 30 pieces of prototype information in descending order of the number of updates.

[0083] Next, the prototype deletion unit 33 deletes the prototype based on a prototype deletion instruction from the operator (S305). The prototype deletion instruction is input by the operator operating the learning device 1 via the input unit 13 by selecting the prototype information displayed on the output unit 14. When the processing of S305 ends, the prototype deletion process ends.

[0084] (Operation and effect of the learning device according to the first embodiment) The learning device 1 autonomously adds prototypes, and therefore can execute learning processing sequentially and continuously for a longer period of time than existing learning devices.

[0085] In addition, the learning device 1 calculates the confidence level R i is the threshold R TH Larger and predicted label L(m j ) does not match the training label L(x), the learning model is not trained, so the prototype vector is not learned by mistake.

[0086] In addition, the learning device 1 calculates the confidence level R i is the threshold R THLarger and predicted label L(m j ) does not match the teacher label L(x), a new prototype is added, allowing us to have multiple prototypes associated with the same label.

[0087] In addition, the learning device 1 calculates the confidence level R i Autonomously learning rate α i Therefore, it is not necessary to repeat the learning process in order to set an appropriate learning rate to improve the learning accuracy, and the learning efficiency is improved.

[0088] In addition, in the learning device 1, the learning rate α i As learning progresses, the confidence level R i decreases as increases, so the confidence R i When is approximately 100%, the learning rate α i becomes approximately 0%, and the learning device 1 can increase the confidence R i High predicted label L(m j ) can be output.

[0089] In addition, the learning device 1 calculates the learning rate α corresponding to all prototypes based on one learning data. i is updated, so that the learning device 1 can obtain the confidence R i High predicted label L(m j ) can be output.

[0090] Furthermore, by compressing prototypes that have been updated less frequently, the learning device 1 can remove prototypes to which incorrect label data has been assigned, thereby reducing the risk of the machine learning model making an incorrect prediction.

[0091] Furthermore, the learning device 1 removes unnecessary prototypes by deleting prototypes that have been updated infrequently, and automatically outputs prototypes with low reliability, thereby reducing the cost of removing unnecessary prototypes.

[0092] (Configuration and Functions of Learning Device According to Second Embodiment) FIG. 12 is a block diagram of a learning device according to the second embodiment.

[0093] Learning device 2 differs from learning device 1 in that it has a processing unit 40 instead of processing unit 20. Processing unit 40 differs from processing unit 20 in that it has an image generation unit 41, a prototype information output unit 42, and a prototype deletion unit 43 instead of the prototype extraction unit 31, the prototype information output unit 32, and the prototype deletion unit 33. The configurations and functions of the components of learning device 2 other than image generation unit 41, prototype information output unit 42, and prototype deletion unit 43 are the same as the configurations and functions of the components of learning device 1 with the same reference numerals, and therefore will not be described in detail here.

[0094] (Prototype deletion process by learning device 2) Fig. 13 is a flowchart showing the prototype deletion process by the learning device 2. The learning process shown in Fig. 13 is executed mainly by the processing unit 40 in cooperation with each element of the learning device 2, based on a control program stored in advance in the storage unit 12.

[0095] First, the image generating unit 41 acquires a prototype deletion instruction (S401). The prototype deletion instruction is input via the input unit 13 by the operator operating the learning device 2, for example, in the same manner as in the process of S301.

[0096] Next, the image generation unit 41 refers to the prototype table 17 and generates prototype image data representing a prototype image that displays each prototype in a manner corresponding to the number of updates (S402). The image generation unit 41 generates prototype image data representing a prototype image in which each prototype has an area corresponding to the number of updates of the prototype, for example. Note that in the prototype image, each prototype may have a side length corresponding to the number of updates of the prototype, instead of having an area corresponding to the number of updates of the prototype, or may be displayed in a color corresponding to the number of updates of the prototype. The image generation unit 41 stores the generated prototype image data in the storage unit 12.

[0097] Next, prototype information output unit 32 outputs the prototype image data generated by image generation unit 41 to output unit 14 (S403). Output unit 14 displays the input prototype image data.

[0098] FIG. 14(a) is a first example of a prototype image displayed on the output unit 14, and FIG. 14(b) is a second example of a prototype image displayed on the output unit 14. FIG. 14(c) is a third example of a prototype image displayed on the output unit 14, and FIG. 14(d) is a fourth example of a prototype image displayed on the output unit 14. In FIGS. 14(a) to 14(d), training data images 301 to 304 are images containing training data used by the learning device 2, and prototype images 311 to 314 are images containing prototypes generated by the learning device. In each of the prototype images 311 to 314 shown in FIGS. 14(a) to 14(d), each prototype has an area corresponding to the number of times the prototype has been updated.

[0099] As shown in Figure 14(a), prototype image 311 shows a prototype image corresponding to prototype image data generated from two sets of training data arranged in a ring shape. In Figure 14(a), circles indicate prototypes corresponding to the first training label, and triangles indicate prototypes corresponding to the second training label. In prototype image 311, the prototypes are arranged in a ring shape, so the number of prototype updates is approximately the same and the area of ​​the prototypes is approximately the same.

[0100] As shown in Figure 14(b), prototype image 312 represents a prototype image corresponding to prototype image data generated from two sets of training data arranged in a crescent shape. In Figure 14(b), circles represent prototypes corresponding to the first training label, and triangles represent prototypes corresponding to the second training label. In prototype image 312, the number of updates for the prototypes corresponding to the first training label are approximately the same, and the areas of the prototypes are approximately the same. On the other hand, the number of updates for the prototypes corresponding to the second training label increases from the edge to the center, and the areas of the prototypes closer to the center are larger than the areas of the prototypes at the edges.

[0101] As shown in Figure 14(c), prototype image 313 shows a prototype image corresponding to prototype image data generated from three sets of training data spaced apart. In Figure 14(c), circles indicate prototypes corresponding to the first training label, triangles indicate prototypes corresponding to the second training label, and squares indicate prototypes corresponding to the third training label. In prototype image 313, one of the prototypes corresponding to the second training label has been updated less frequently than the other prototypes, and has a smaller area. The number of updates for all prototypes other than one corresponding to the second training label are approximately the same, and the areas of all prototypes other than one corresponding to the second training label are approximately the same.

[0102] As shown in FIG. 14(d), prototype image 314 shows a prototype image corresponding to prototype image data generated from four sets of training data. In FIG. 14(d), circles indicate prototypes corresponding to the first training label, triangles indicate prototypes corresponding to the second training label, squares indicate prototypes corresponding to the third training label, and diamonds indicate prototypes corresponding to the fourth training label. The prototypes corresponding to the first training label and the fourth training label are positioned closely together. On the other hand, the prototypes corresponding to the first training label and the fourth training label are positioned farther apart from the prototypes corresponding to the second training label and the prototype corresponding to the third training label. In prototype image 314, the prototypes positioned between the closely positioned prototypes corresponding to the first training label and the fourth training label have been updated less frequently and have a smaller area than the other prototypes.

[0103] Next, the prototype deletion unit 43 deletes the prototype based on a prototype deletion instruction from the operator (S404). The prototype deletion instruction is input by the operator operating the learning device 2 via the input unit 13 by selecting the prototype information displayed on the output unit 14. When the processing of S404 ends, the prototype deletion process ends.

[0104] (Operation and effect of the learning device according to the second embodiment) The learning device 2 displays prototype images that display each prototype in a format according to the number of updates, making it easy for the operator to visually identify prototypes that have been updated less frequently, facilitating the operator's work of removing prototypes.

[0105] (Configuration and Functions of Learning Device According to Third Embodiment) FIG. 15 is a block diagram of a learning device according to the third embodiment.

[0106] Learning device 3 differs from learning device 1 in that it has processing unit 50 instead of processing unit 20. Processing unit 50 differs from processing unit 20 in that it has prototype deletion unit 51 and deletion determination unit 52 instead of prototype extraction unit 31, prototype information output unit 32, and prototype deletion unit 33. The configurations and functions of the components of learning device 3 other than prototype deletion unit 51 and deletion determination unit 52 are the same as the configurations and functions of the components of learning device 1 with the same reference numerals, and therefore will not be described in detail here.

[0107] (Prototype deletion process by learning device 3) Fig. 16 is a flowchart showing the prototype deletion process by the learning device 3. The learning process shown in Fig. 16 is executed mainly by the processing unit 50 in cooperation with each element of the learning device 3, based on a control program stored in advance in the storage unit 12.

[0108] First, the prototype deletion unit 51 acquires a prototype deletion instruction (S501). As in the process of S301, the prototype deletion instruction is input via the input unit 13 by the operator operating the learning device 3, for example.

[0109] Next, the prototype deletion unit 51 refers to the prototype table 17 and deletes the prototype with the least number of updates (S502). The prototype deletion unit 51 compares the number of updates corresponding to the update count information in the prototype table 17, and deletes from the prototype table 17 the prototype with the least number of updates corresponding to the update count information.

[0110] Next, the deletion determination unit 52 stores the certainty factor by referring to the certainty factor table 18 (S503). The deletion determination unit 52 stores the certainty factor stored in the certainty factor table 18 in the storage unit 12 as the pre-update certainty factor.

[0111] Next, the learning data acquisition unit 21 to the update count increase unit 28 execute the learning process (S504). The learning process executed by the learning data acquisition unit 21 to the update count increase unit 28 has been described with reference to FIG. 7, so a detailed description thereof will be omitted here.

[0112] Next, the deletion determination unit 52 determines whether or not to terminate the deletion of prototypes by the prototype deletion unit 51 based on the pre-update confidence levels stored in the process of S503 and the confidence levels stored in the confidence level table 18 (S505). The deletion determination unit 52 determines to terminate the deletion of prototypes by the prototype deletion unit 51 when at least one of the confidence levels stored in the confidence level table 18 has increased compared to the corresponding confidence level of the pre-update confidence levels stored in the process of S502 (YES in S505). On the other hand, the deletion determination unit 52 determines not to terminate the deletion of prototypes by the prototype deletion unit 51 when at least one of the confidence levels stored in the confidence level table 18 has decreased compared to the corresponding confidence level of the pre-update confidence levels stored in the process of S502 (NO in S505).

[0113] If the deletion determination unit 52 determines not to end the deletion of the prototype (S505-NO), the process returns to S501. Thereafter, the processes of S502 to S505 are repeated until the deletion determination unit 52 determines to end the deletion of the prototype (S505-YES). If the deletion determination unit 52 determines to end the deletion of the prototype (S505-YES), the prototype deletion process ends.

[0114] (Operational effects of the learning device according to the third embodiment) The learning device 3 sequentially deletes the prototype with the least number of updates while determining whether to terminate the deletion of prototypes based on the confidence level, thereby enabling the learning device 3 to autonomously delete prototypes until the number of prototypes reaches an appropriate number.

[0115] (Modification of the learning device according to the embodiment) Although the learning devices 1 to 3 use an LVQ neural network as a learning model, the learning device according to the embodiment may use a learning model that allows online machine learning using a learning rate as the learning model. Furthermore, the learning device 1 stores the learning model 16 in the memory unit 12, but the learning device according to the embodiment may use a learning model stored in a server that is communicably connected via a communication network such as the Internet or a LAN, rather than storing the learning model in the memory unit.

[0116] Furthermore, the learning devices 1 to 3 calculate the confidence R i However, the learning device according to the embodiment calculates at least the teacher label L(x) and the predicted label L(m j ) and the co-occurrence frequency can be changed based on the degree of agreement between them.

[0117] For example, the confidence level R i may be calculated using equation (4) below:

[0118]

number

[0119] Using equation (4), the confidence level R i When calculating, the first co-occurrence frequency a i indicates the number of times the predicted label matches the teacher label, and the second co-occurrence frequency b i may indicate the number of times the predicted label matches the teacher label.

[0120] Also, the confidence level R i may be calculated using any of the following equations (5) to (7).

[0121]

number

[0122]

number

[0123]

number

[0124] Furthermore, the learning devices 1 to 3 use the formula (2) to calculate the learning rate α i However, the learning device according to the embodiment calculates the confidence R i As the learning rate α increases, i The learning rate may be calculated so that the learning rate decreases as the ratio of the first co-occurrence frequency to the total frequency of the first co-occurrence frequency and the second co-occurrence frequency increases. For example, the learning device 1 may update the learning rate so that the learning rate decreases as the ratio of the first co-occurrence frequency to the total frequency of the first co-occurrence frequency, the second co-occurrence frequency, and the third co-occurrence frequency increases. Furthermore, the learning rate may be updated so that the learning rate decreases as the ratio of the sum of the first co-occurrence frequency and the fourth co-occurrence frequency to the total frequency of the first co-occurrence frequency, the second co-occurrence frequency, the third co-occurrence frequency, and the fourth co-occurrence frequency increases.

[0125] In the learning devices 1 to 3, the first co-occurrence frequency a i , second co-occurrence frequency b i , the third co-occurrence frequency c i and the fourth co-occurrence frequency d i However, in the learning device according to the embodiment, it is sufficient to define it as a parameter indicating the frequency with which each event occurs.

[0126] Furthermore, although learning devices 1 to 3 are devices that add and delete prototypes, they may also be used as learning data compression devices that compress the learning data to be used. When learning devices 1 to 3 are used as learning data compression devices, prototype adding unit 23 functions as a learning data adding unit that determines whether or not to add learning data to memory unit 12 as learning data included in the compressed learning data, based on the confidence factor, teacher label, and predicted label.

[0127] By using the learning devices 1 to 3 as learning data compression devices, it is possible to compress learning data for generating a learning model with high prediction accuracy. For example, the learning data shown in each of Figures 14(a) to 14(d) is compressed to 1.9% to 6.0% by the learning process shown in Figures 7 and 8. Furthermore, the learning data shown in each of Figures 10(a) to 10(d) and 11(a) to 11(b) is compressed to 9.9% to 29.1% by the learning process shown in Figures 7 and 8.

[0128] It was confirmed that the training model trained using the training data compressed by the training process shown in Figures 7 and 8 had the same prediction accuracy as the training model trained using the training data before compression. Regarding the training data shown in Figures 14(a) to 14(d), the training data shown in Figure 14(b) showed a decrease in prediction accuracy compared to before compression, while the training data shown in Figures 14(a), 14(c), and 14(d) maintained the same prediction accuracy as before compression. Furthermore, the training data shown in Figures 10(a) to 10(d) and Figures 11(a) to 11(b), respectively, showed a decrease in prediction accuracy of approximately 0.4% to 3.2% compared to before compression.

[0129] The learning devices 1 to 3 used as training data compression devices can dynamically remove unnecessary training data by adding the training data to the storage unit 12 as training data included in the compressed training data based on the confidence level, the teacher label, and the predicted label. Furthermore, when training data is added, the learning devices 1 to 3 check the number of updates of the prototype most similar to the training data to be added, thereby automatically outputting low-reliability training data and reducing the cost of removing unnecessary training data.

[0130] Furthermore, the learning device 3 determines to end the deletion of a prototype when the confidence level after the prototype is deleted is higher than the confidence level before the prototype was deleted. However, the learning device according to the embodiment may determine whether to end the deletion of a prototype based on the confidence level. For example, the learning device according to the embodiment may determine to end the deletion of a prototype when the confidence level drops to a predetermined threshold value. [Explanation of symbols]

[0131] 1-3 Learning device (learning data compression device) 16 Learning Model 21 Learning data acquisition unit 22 Predictive label acquisition unit 23 Prototype addition unit (learning data addition unit) 24 Co-occurrence frequency change section 25 Confidence calculation section 26 Learning rate calculation unit 27 Prototype feature update unit 28 Update count increase section 31 Prototype Extraction 32, 42 Prototype information output section 33, 43, 51 Prototype deletion section 41 Image generation unit 52 Deletion determination unit

Claims

1. a storage unit that stores a plurality of prototypes, each of which includes a prototype label, a prototype feature corresponding to the prototype label, and update count information indicating the number of times the prototype feature has been updated; a training data acquisition unit that acquires training data including input feature data indicating input feature amounts and truth label data indicating truth labels associated with the input feature amounts; a predicted label acquisition unit that acquires a predicted label output by a learning model to which the input feature amount is input; a prototype feature updating unit that updates the prototype features based on at least a degree of agreement between the teacher label and the predicted label; an update count incrementing unit that increments by one the update count corresponding to the update count information of the prototype whose prototype feature has been updated; a prototype adding unit that determines whether to add a prototype to the storage unit based on a confidence factor indicating the likelihood that the learning model will output the teacher label associated with the input feature as the predicted label in response to input of the input feature, the teacher label, and the predicted label; and a prototype deletion unit that deletes one of the prototypes based on an operator's instruction or the number of updates; A learning device comprising:

2. a prototype extraction unit that extracts a prototype whose update count satisfies a predetermined condition; a prototype information output unit that outputs prototype information indicating the prototypes extracted by the prototype extraction unit; The learning device according to claim 1 , further comprising:

3. The learning device according to claim 2 , wherein the prototype extraction unit extracts the prototypes whose number of updates is less than a predetermined threshold value.

4. The learning device according to claim 2 , wherein the prototype extraction unit extracts a predetermined number of the prototypes in ascending order of the number of updates.

5. an image generating unit that generates prototype image data representing a prototype image that displays each of the prototypes in a manner corresponding to the number of updates; a prototype information output unit that outputs the prototype image data generated by the image generation unit; The learning device according to claim 1 , further comprising:

6. The prototype deletion unit deletes the prototype that has been updated the least number of times, a deletion determination unit that determines whether or not to terminate the deletion of the prototype by the prototype deletion unit based on the confidence level; The learning device according to claim 1 , further comprising:

7. 7. The learning device according to claim 6, wherein the deletion determination unit determines to terminate the deletion of the prototype by the prototype deletion unit when the certainty factor after the prototype is deleted is higher than the certainty factor before the prototype is deleted.

8. 1. A learning method for causing a learning device to learn a learning model, the learning device having a storage unit that stores a plurality of prototypes, each of the prototypes including a prototype label, a prototype feature corresponding to the prototype label, and update count information indicating the number of times the prototype feature is updated, the method comprising: acquiring learning data including input feature data indicating input feature values ​​and truth label data indicating truth labels associated with the input feature values; Obtain a predicted label output by a learning model to which the input feature is input; updating the prototype features based on at least a degree of agreement between the training labels and the predicted labels; Incrementing the update count corresponding to the update count information of the prototype whose prototype feature has been updated by 1; determining whether to add a prototype to a storage unit based on a confidence level indicating the likelihood that the learning model will output the teacher label associated with the input feature as the predicted label in response to the input of the input feature, the teacher label, and the predicted label; deleting one of the prototypes based on the number of updates or in response to an operator's instruction based on the number of updates; A learning method comprising:

9. 1. A control program for causing a learning device to learn a learning model, the control program including a storage unit that stores a plurality of prototypes, each of the prototypes including a prototype label, a prototype feature corresponding to the prototype label, and update count information indicating the number of updates to the prototype feature, the control program comprising: acquiring learning data including input feature data indicating input feature values ​​and truth label data indicating truth labels associated with the input feature values; Obtain a predicted label output by a learning model to which the input feature is input; updating the prototype features based on at least a degree of agreement between the training labels and the predicted labels; Incrementing the update count corresponding to the update count information of the prototype whose prototype feature has been updated by 1; determining whether to add a prototype to a storage unit based on a confidence level indicating the likelihood that the learning model will output the teacher label associated with the input feature as the predicted label in response to the input of the input feature, the teacher label, and the predicted label; deleting one of the prototypes based on the number of updates or in response to an operator's instruction based on the number of updates; A learning program that causes the learning device to execute a process.

10. a storage unit that stores a prototype label and a plurality of prototypes, each of which includes a prototype feature corresponding to the prototype label; a training data acquisition unit that acquires training data including input feature data indicating input feature amounts and truth label data indicating truth labels associated with the input feature amounts; a predicted label acquisition unit that acquires a predicted label output by a learning model to which the input feature amount is input; a learning data adding unit that determines whether to add the learning data to the storage unit based on a confidence factor indicating a probability that the learning model will output the teacher label associated with the input feature as the predicted label in response to input of the input feature, the teacher label, and the predicted label; and a prototype deletion unit that deletes one of the prototypes based on an operator's instruction or the number of updates; A training data compression device comprising:

Images