Learning device, learning method, and learning program

Abstract

This learning device trains an AI model using a dataset in which each piece of data is given a mnemonic code corresponding to a label of that piece of data. Thereafter, the learning device uses the mnemonic code for each label to calculate the diagonal component of the F matrix (pseudo Fisher information matrix) with respect to each label learned by the AI model. Next, the learning device determines noise to impose on a parameter of the AI model on the basis of the value of the diagonal component of the pseudo F matrix with respect to a label to be forgotten and the value of the diagonal component of the F matrix with respect to a label not to be forgotten. The learning device then updates the parameter by imposing the determined noise on the parameter of the AI model.

Description

[Technical Field]

[0001] The present invention relates to a learning device, a learning method, and a learning program that cause an AI model to forget learned labels. [Background technology]

[0002] To protect privacy and avoid data leaks, there is a challenge known as selective forgetting, which involves making an AI model forget desired tasks that it has learned through machine learning in the past. [Prior art documents] [Non-patent literature]

[0003] [Non-Patent Document 1] Tomoya Yamashita et al., Research on AI knowledge forgetting based on the importance of model parameters, Institute of Electronics, Information and Communication Engineers, IEICE Technical Report NC2022-3, IBISML2022-3 (2022-06) Summary of the Invention [Problem to be solved by the invention]

[0004] However, existing technologies have not been able to cause an AI model to forget data that belongs to a specified label from a dataset that the AI ​​model has learned. Therefore, an object of the present invention is to solve the above problem and cause an AI model to forget data that belongs to a specified label. [Means for solving the problem]

[0005] In order to solve the above-mentioned problems, the present invention is characterized by comprising an AI model learning unit that learns an AI model using a dataset in which each piece of data is assigned a code corresponding to the label of the data; a calculation unit that calculates a value indicating the importance of the AI ​​model parameters for each label learned by the AI ​​model using the code corresponding to the label; a noise determination unit that determines noise to be added to the parameters of the AI ​​model based on the value indicating the importance of the AI ​​model parameters for labels to be forgotten and the value indicating the importance of the AI ​​model parameters for labels not to be forgotten; and a parameter update unit that updates the parameters of the AI ​​model by adding the determined noise to the parameters of the AI ​​model. [Effects of the Invention]

[0006] According to the present invention, data that belongs to a specified label can be forgotten from an AI model. [Brief explanation of the drawings]

[0007] [Figure 1] Figure 1 is a diagram illustrating an example of AI model learning using EWC. [Figure 2] FIG. 2 is a diagram illustrating an overview of the learning device. [Figure 3] FIG. 3 is a diagram illustrating an example of the configuration of a learning device. [Figure 4] FIG. 4 is a flowchart showing an example of a procedure for label forgetting processing. [Figure 5] FIG. 5 is a flowchart illustrating an example of a procedure for prediction processing of input data. [Figure 6] FIG. 6 shows the results of an evaluation experiment of an AI model in which label forgetting processing was performed by a learning device. [Figure 7] FIG. 7 shows the results of an evaluation experiment of an AI model in which label forgetting processing was performed by a learning device. [Figure 8] FIG. 8 is a diagram illustrating an example of the configuration of a computer that executes a learning program. DETAILED DESCRIPTION OF THE INVENTION

[0008] Hereinafter, a mode (embodiment) for carrying out the present invention will be described with reference to the drawings. The present invention is not limited to this embodiment. First, a prerequisite technology used in the learning device of this embodiment will be described.

[0009] [Catastrophic Forgetting] Catastrophic forgetting is a phenomenon in which an AI model forgets past tasks when it is trained on multiple tasks consecutively. Continual learning, described below, aims to prevent this catastrophic forgetting while learning new tasks. Furthermore, existing technologies for learning with selective forgetting utilize catastrophic forgetting to enable an AI model to forget desired tasks.

[0010] [Continual Learning] Continual learning is an AI research field that aims to achieve continuous learning of multiple tasks. A key challenge in continual learning is to learn new tasks without losing accuracy on previously learned tasks (avoiding catastrophic forgetting).

[0011] Continual Learning defines a set of datasets used for learning as D={D1,D2,…,D K}. Then, when learning an AI model, D K Continuously learn until

[0012] The purpose of this continual learning is to develop an AI model (f θ :X→Y).

[0013] [EWC] Elastic Weight Consolidation (EWC) is a continual learning method that avoids catastrophic forgetting by adding a regularization term to the loss function used when training an AI model.

[0014] The regularization term added in EWC penalizes fluctuations in parameters of the AI ​​model that are important for solving previously learned tasks, allowing the AI ​​model to learn new tasks without degrading its performance on previously learned tasks.

[0015] Figure 1 shows how learning progresses toward the intersection of parameters suitable for previously learned tasks and parameters suitable for a new task due to the effect of the EWC regularization term (details to be described later). Note that this example explains the case where an AI model learns task A and then learns task B. L2 is the parameter θ for task A. A * "no penalty" indicates learning of parameters for task B with a penalty of not changing them as much as possible from . Also, "no penalty" indicates learning of parameters for task B without a penalty.

[0016] In EWC, the loss function used when learning Task B after learning Task A is shown in Equation (1) below.

[0017]

number

[0018] L in Equation (1) B (θ) is the loss function for task B, and θ A is the parameter vector of the AI ​​model immediately after learning task A.

[0019] In this formula (1), a regularization term is added to the general loss function for learning Task B. This regularization term ensures that the parameters of the AI ​​model, which are important for solving Task A, do not fluctuate significantly due to learning Task B. Specifically, in formula (1), the fluctuation of the parameters due to learning Task B ((θ i -θ A,i ) 2 ) for which the Fisher information matrix (diagonal components) of the parameters is F i The weighting is done according to the importance of the parameter. Note that λ in Equation (1) is a hyperparameter.

[0020] [Learning with Selective Forgetting (LSF)] Learning with Selective Forgetting (LSF) is an AI technology challenge that aims to continuously learn multiple tasks while forgetting knowledge of desired tasks from previously learned tasks.

[0021] The set of datasets used for learning is defined in the same way as for Continual Learning. The following two sets of datasets are defined for learning.

[0022] Preservation Set: A set of classes to be memorized after completing task learning Deletion Set: A set of classes to be forgotten after the task has been trained.

[0023] In LSF, the goal is to retain knowledge of the classes in the Preservation Set after learning each task, while forgetting the classes in the Deletion Set.

[0024] [Units to be forgotten] The data units that the AI ​​model forgets are classified into the following three types:

[0025] (1) Dataset-level forgetting An AI model learns different datasets using continual learning techniques, and then forgets about one dataset. (2) Label-based forgetting The process by which an AI model forgets data that belongs to a specific label from the dataset it has learned. (3) Forgetting data units Forgetting a certain piece of data from a dataset that an AI model has learned.

[0026] The learning device 10 of this embodiment performs (2) label-based forgetting among the above.

[0027] [Overview of the learning device] Next, an overview of the learning device 10 of this embodiment will be described using Fig. 2. The system includes, for example, a learning device 10 that performs learning of an AI model, and a data acquisition device 20. The AI ​​model is, for example, a model that outputs a predicted value of input data acquired from the data acquisition device 20.

[0028] In conventional technologies, it is thought that the reason why it was difficult for an AI model to forget data on a label-by-label basis is that the data (e.g., image data) included in the training data has a high degree of similarity between different labels. Therefore, in conventional technologies, it is possible that the parameters used by the AI ​​model are not sufficiently separated into parameters that are important for each label. Therefore, the learning device 10 of this embodiment trains the AI ​​model using a dataset containing a mnemonic code for each label. This reduces the similarity between labels calculated by the AI ​​model.

[0029] The learning device 10 uses a data set to learn the AI ​​model. This data set is, for example, a data set of labeled image data, with noise (mnemonic code) for each label added to each piece of image data.

[0030] The learning device 10 can reduce the similarity between image data of different labels in the AI ​​model by training the AI ​​model using a dataset of image data carrying a mnemonic code for each label, thereby making it easier for the learning device 10 to forget on a label-by-label basis in the AI ​​model.

[0031] Thereafter, the learning device 10 causes the label designated by the user to be forgotten among the labels learned by the AI ​​model. For example, the learning device 10 first calculates the diagonal elements of a pseudo Fisher information matrix (F matrix for each label) of the parameters of the AI ​​model for each label learned by the AI ​​model according to the implementation of EWC. Note that when calculating the pseudo F matrix, a mnemonic code corresponding to each label is used.

[0032] Then, the learning device 10 uses the diagonal elements of the F matrix for each calculated label to determine noise to be added to the parameters of the AI ​​model in accordance with the label to be forgotten.

[0033] This allows the learning device 10 to determine parameter noise that is effective in forgetting the label specified by the user.The learning device 10 then applies the determined noise to the parameters of the AI ​​model and updates the parameters of the AI ​​model.As a result, the learning device 10 can cause the AI ​​model to forget the label specified by the user.

[0034] [Example of learning device configuration] Next, an example of the configuration of the learning device 10 will be described with reference to Fig. 3. The learning device 10 includes, for example, an input unit 11, an output unit 12, a communication control unit 13, a storage unit 14, and a control unit 15.

[0035] The input unit 11 is an interface that accepts input of various data. For example, the input unit 11 accepts input of a data set used for learning the AI ​​model, an instruction input of labels to be forgotten, and input of data to be predicted by the AI ​​model. The output unit 12 is an interface that outputs various data. For example, the output unit 12 outputs the label of data predicted by the control unit 15.

[0036] The communication control unit 13 is realized by a NIC (Network Interface Card) or the like, and controls communication between an external device such as a server via a network and the control unit 15. For example, the communication control unit 13 controls communication between the control unit 15 and a data acquisition device 20 (see FIG. 2 ) or the like.

[0037] The storage unit 14 is realized by a semiconductor memory element such as a random access memory (RAM) or a flash memory, or a storage device such as a hard disk or an optical disk, and stores parameters of the AI ​​model.

[0038] The control unit 15 is realized using, for example, a CPU (Central Processing Unit) or the like, and executes a processing program stored in the storage unit 14. As a result, the control unit 15 functions as an acquisition unit 15a, a learning unit 15b, and a prediction unit 15c illustrated in FIG.

[0039] The acquisition unit 15a acquires, for example, a data set used for learning an AI model, labels to be forgotten, and the like via the input unit 11 or the communication control unit 13.

[0040] The learning unit 15b learns an AI model and performs a forgetting process on a label specified for the learned AI model. The learning unit 15b includes an AI model learning unit 150, a Fisher information matrix calculation unit (calculation unit) 151, a noise determination unit 152, and a parameter update unit 153.

[0041] The AI ​​model learning unit 150 uses a data set to learn the AI ​​model. This data set is, for example, a data set in which each piece of data (e.g., image data) is provided with a mnemonic code corresponding to the label of the data.

[0042] The Fisher information matrix calculation unit 151 calculates a value indicating the importance of the parameters of the AI ​​model for each label learned by the AI ​​model. For example, the Fisher information matrix calculation unit 151 calculates the diagonal elements of a pseudo-F matrix for each label of the AI ​​model learned by the AI ​​model learning unit 150.

[0043] For example, a general Fisher information matrix for label j uses a value obtained by second-order differentiation of equation (2) with parameter w, but the Fisher information matrix calculation unit 151 uses a pseudo Fisher information matrix (pseudo-F matrix) using a value obtained by squaring a value obtained by first-order differentiation of loss L with parameter w. Note that when the Fisher information matrix calculation unit 151 calculates the diagonal elements of the pseudo-F matrix for each label, it uses a mnemonic code corresponding to each label.

[0044]

number

[0045] The noise determination unit 152 determines the noise to be added to the parameters of the AI ​​model based on the diagonal elements of the pseudo-F matrix for each label calculated by the Fisher information matrix calculation unit 151. For example, the noise determination unit 152 determines the noise to be added to the parameters of the AI ​​model based on the values ​​of the diagonal elements of the pseudo-F matrix for labels to be forgotten and the values ​​of the diagonal elements of the pseudo-F matrix for labels not to be forgotten.

[0046] For example, the noise determination unit 152 determines the noise to be added to the parameters of the AI ​​model based on the ratio of the diagonal elements of the pseudo-F matrix for the label j to be forgotten (importance for the label to be forgotten) to the diagonal elements of the pseudo-F matrix for the labels not to be forgotten (importance for the label not to be forgotten) using the following equation (3):

[0047]

number

[0048] Note that m in formula (3) is the number of labels. The noise determination unit 152 uses formula (3) to add large noise to parameters that are important to label j, and to prevent large noise from being added to parameters that are important to labels other than label j.

[0049] The parameter update unit 153 updates the parameters of the AI ​​model by adding the noise determined by the noise determination unit 152 to the parameters. For example, the parameter update unit 153 updates the parameters w of the AI ​​model based on the following equation (4): i The parameters are updated by adding noise to

[0050]

number

[0051] In equation (4), λ1 is a parameter for cutting off the noise added to the parameters of each layer of the AI ​​model. λ2 is a parameter for normalizing the noise added to the parameters of each layer of the AI ​​model. For example, λ1=10 -7 , λ2=1.0.

[0052] Here, as shown in equation (4), when adding noise to the parameters of the AI ​​model, the parameter update unit 153 either adds noise to the parameters or subtracts noise from the parameters.Whether the parameter update unit 153 adopts adding noise to the parameters or subtracting noise from the parameters is determined, for example, as follows.

[0053] For example, for a specified label j, the parameter update unit 153 performs two operations: adding noise to the parameters of the trained AI model, and subtracting noise from the parameters, and determines which is better from the perspective of forgetting. For example, for a specified label j, the parameter update unit 153 determines which operation results in a decrease in the prediction accuracy of the AI ​​model.

[0054] Then, the parameter update unit 153 updates the parameters by adding noise to the parameters of the AI ​​model using a method determined to be superior from the perspective of forgetting (a method of adding noise to the parameters or a method of subtracting noise from the parameters).

[0055] The prediction unit 15c predicts (identifies) the label of the input data using the AI ​​model after the parameter update. For example, the prediction unit 15c calculates the probability of each label of the input data using the AI ​​model after the parameter update, and outputs the label with the highest probability.

[0056] [Label forgetting process] Next, an example of the procedure for forgetting labels of an AI model by the learning device 10 will be described with reference to Fig. 4. The process shown in Fig. 4 starts, for example, when a label to be forgotten is selected by the user and an operation input is made to instruct the start of label forgetting processing.

[0057] First, the AI ​​model learning unit 150 learns the AI ​​model using a data set in which a mnemonic code is added to image data for each label (S10).

[0058] Next, the Fisher information matrix calculation unit 151 calculates a pseudo Fisher information matrix (pseudo F matrix) for each label learned by the AI ​​model using the mnemonic code corresponding to each label, and acquires the diagonal elements of the matrix (S11).

[0059] Next, the noise determination unit 152 determines noise for the parameters of the AI ​​model using the diagonal elements of the pseudo F matrix for each label obtained in S11 (S12). For example, the noise determination unit 152 determines noise to be added to the parameters of the AI ​​model based on the values ​​of the diagonal elements of the pseudo F matrix for the label j to be forgotten and the values ​​of the diagonal elements of the F matrix for the labels not to be forgotten.

[0060] Thereafter, the parameter update unit 153 updates the parameters of the AI ​​model by adding the noise determined in S12 to the parameters of the AI ​​model (by adding to or subtracting from the parameters) (S13).

[0061] For example, for label j designated as a label to be forgotten, the parameter update unit 153 adds noise to the parameters of the learned AI model and subtracts noise from the parameters, and determines which is better from the perspective of forgetting. Then, the parameter update unit 153 updates the parameters by adding noise to the parameters of the AI ​​model using the method determined to be better from the perspective of forgetting (adding noise to the parameters or subtracting noise from the parameters).

[0062] [Prediction processing] Next, an example of input data label prediction processing by the learning device 10 will be described with reference to Fig. 5. The processing shown in Fig. 5 starts, for example, when an operation input is made to instruct the start of input data label prediction processing.

[0063] First, the acquisition unit 15a acquires data for which a label is to be predicted (S21). Next, the prediction unit 15c predicts the label of the data acquired in S21 using an AI model whose parameters have been updated by the learning unit 15b (S22). For example, the prediction unit 15c calculates p(x') of the data x' acquired in S21 using the AI ​​model after the parameter update, and outputs the label with the highest probability.

[0064] [experiment] Experiments 1 and 2 were conducted to confirm that the AI ​​model can forget the label specified by the user using the learning device 10 of this embodiment. The results of Experiments 1 and 2 are shown below.

[0065] [Experiment 1] In Experiment 1, first, the AI ​​model is trained on MNIST. After that, it is confirmed that the AI ​​model can forget the specified labels using the learning device 10. The AI ​​model used in Experiment 1 is a three-layer AI model with an input layer of 28 x 28 units, a hidden layer of 1000 units, and an output layer of 10 units. In Experiment 1, when learning (training) the AI ​​model, MNIST training data and MNIST training data with mnemonic codes added were used. In addition, when testing the AI ​​model after training, MNIST test data was used.

[0066] The learning results for MNIST are shown in Table 601 in Figure 6. As shown in Table 601, it was confirmed that by using EWC, the AI ​​model was able to classify the MNIST labels "0" to "9" with an accuracy of over 86%.

[0067] Next, Table 602 in Figure 6 shows the results of the learning device 10 performing a forgetting process for the label "0" on an AI model that has learned MNIST. The values ​​shown in the first row of Table 602 are the accuracy when noise is added to the parameters of the AI ​​model, and the values ​​shown in the second row of Table 602 are the accuracy when noise is subtracted from the parameters of the AI ​​model. Note that when the learning device 10 adds noise to the parameters of the AI ​​model and when it subtracts noise from the parameters, λ2 = 1 in the above-mentioned equation (3) is used.

[0068] As shown in Table 602, when noise was added to the parameters of the AI ​​model, the accuracy of the label “0” became “5.41%”, confirming the forgetting effect of the label “0”.

[0069] [Experiment 2] In Experiment 2, we conducted the same experiment as Experiment 1 using a dataset called CIFAR10. The AI ​​model used in Experiment 2 was ResNet18. When training the AI ​​model, we used CIFAR10 training data and CIFAR10 training data with mnemonic codes added. In addition, when testing the trained AI model, we used CIFAR10 test data.

[0070] The learning results of CIFAR10 are shown in Table 701 in Figure 7. As shown in Table 701, it was confirmed that by using EWC, the AI ​​model can classify the labels "0" to "9" with an accuracy of over 70%.

[0071] Next, Table 702 in Figure 7 shows the results of the learning device 10 performing a forgetting process for the label "0" on an AI model that has learned CIFAR10. The values ​​shown in the first row of Table 702 are the accuracy when noise is added to the parameters of the AI ​​model, and the values ​​shown in the second row of Table 702 are the accuracy when noise is subtracted from the parameters of the AI ​​model. Note that when the learning device 10 adds noise to the parameters of the AI ​​model and when it subtracts noise from the parameters, λ2 = 0.9 in the above-mentioned formula (3) is used.

[0072] As shown in Table 702, in an experiment using CIFAR10, when noise was subtracted from the parameters of the AI ​​model, the accuracy of the label "0" was "0.2%", confirming the forgetting effect of the label "0".

[0073] From the above experiments 1 and 2, it was confirmed that the learning device 10 of this embodiment enables the AI ​​model to forget the label specified by the user.

[0074] [System configuration, etc.] Furthermore, the components of each unit shown in the figure are conceptual functional units and do not necessarily have to be physically configured as shown. In other words, the specific form of distribution and integration of each device is not limited to that shown, and all or part of them can be functionally or physically distributed and integrated in any unit depending on various loads, usage conditions, etc. Furthermore, all or any part of the processing functions performed by each device can be realized by a CPU and a program executed by the CPU, or can be realized as hardware using wired logic.

[0075] Furthermore, among the processes described in the above embodiments, all or part of the processes described as being performed automatically can be performed manually, or all or part of the processes described as being performed manually can be performed automatically using a known method.In addition, the information including the processing procedures, control procedures, specific names, various data and parameters shown in the above documents and drawings can be changed as desired unless otherwise specified.

[0076] [program] The learning device 10 can be implemented by installing a program (learning program) as package software or online software on a desired computer. For example, by running the program on an information processing device, the information processing device can function as learning device 10. The information processing device referred to here includes mobile communication terminals such as smartphones, mobile phones, and PHS (Personal Handyphone Systems), as well as terminals such as PDAs (Personal Digital Assistants).

[0077] 8 is a diagram showing an example of a computer that executes a learning program. The computer 1000 includes, for example, a memory 1010 and a CPU 1020. The computer 1000 also includes a hard disk drive interface 1030, a disk drive interface 1040, a serial port interface 1050, a video adapter 1060, and a network interface 1070. These components are connected by a bus 1080.

[0078] The memory 1010 includes a ROM (Read Only Memory) 1011 and a RAM (Random Access Memory) 1012. The ROM 1011 stores a boot program such as a BIOS (Basic Input Output System). The hard disk drive interface 1030 is connected to a hard disk drive 1090. The disk drive interface 1040 is connected to a disk drive 1100. A removable storage medium such as a magnetic disk or optical disk is inserted into the disk drive 1100. The serial port interface 1050 is connected to a mouse 1110 and a keyboard 1120, for example. The video adapter 1060 is connected to a display 1130, for example.

[0079] The hard disk drive 1090 stores, for example, an OS 1091, an application program 1092, a program module 1093, and program data 1094. That is, the programs that define the processes executed by the learning device 10 are implemented as program modules 1093 in which computer-executable code is written. The program modules 1093 are stored, for example, on the hard disk drive 1090. For example, a program module 1093 for executing processes similar to those of the functional configuration of the learning device 10 is stored on the hard disk drive 1090. The hard disk drive 1090 may be replaced by an SSD (Solid State Drive).

[0080] Data used in the processing of the above-described embodiment is stored as program data 1094 in, for example, the memory 1010 or the hard disk drive 1090. The CPU 1020 then reads the program module 1093 or the program data 1094 stored in the memory 1010 or the hard disk drive 1090 into the RAM 1012 as needed and executes them.

[0081] The program module 1093 and program data 1094 are not limited to being stored in the hard disk drive 1090, but may also be stored in, for example, a removable storage medium and read by the CPU 1020 via the disk drive 1100 or the like. Alternatively, the program module 1093 and program data 1094 may be stored in another computer connected via a network (such as a LAN (Local Area Network) or WAN (Wide Area Network)). The program module 1093 and program data 1094 may then be read by the CPU 1020 from the other computer via the network interface 1070. [Explanation of symbols]

[0082] 10 Learning Device 11 Input section 12 Output section 13 Communication control section 14 Storage section 15 Control Unit 15a Acquisition part 15b Learning Department 15c Prediction Section 20 Data acquisition device 150 AI Model Learning Department 151 Fisher Information Matrix Calculation Unit 152 Noise determination unit 153 Parameter Update Unit

Claims

1. an AI model learning unit that learns an AI model using a dataset in which each piece of data has a code corresponding to the label of the data; a calculation unit that calculates a value indicating the importance of a parameter of the AI ​​model for each label learned by the AI ​​model using the code corresponding to the label; a noise determination unit that determines noise to be added to the parameters of the AI ​​model based on a value indicating the importance of the parameters of the AI ​​model related to the label to be forgotten and a value indicating the importance of the parameters of the AI ​​model related to the label not to be forgotten; a parameter update unit that updates the parameters of the AI ​​model by adding the determined noise to the parameters of the AI ​​model; A learning device comprising:

2. The code Mnemonic Code corresponding to the label 2. The learning device according to claim 1, wherein:

3. The calculation unit Calculating the diagonal elements of the Fisher information matrix of the parameters of the AI ​​model for each label learned by the AI ​​model as a value indicating the importance of the parameters of the AI ​​model 2. The learning device according to claim 1 .

4. The noise determination unit The noise to be added to the parameters of the AI ​​model is determined based on the ratio of a value indicating the importance of the parameters of the AI ​​model for the labels to be forgotten to a value indicating the importance of the parameters of the AI ​​model for the labels not to be forgotten.

2. The learning device according to claim 1 .

5. The parameter update unit updating the parameters of the AI ​​model by adding or subtracting the determined noise to or from the parameters of the AI ​​model; 2. The learning device according to claim 1 .

6. A prediction unit that outputs a predicted value of the label of the input data using the AI ​​model with the updated parameters. The learning device according to claim 1 , further comprising:

7. A learning method executed by a learning device, comprising: A step of training an AI model using a dataset in which each piece of data is assigned a code corresponding to the label of the data; calculating a value indicating the importance of a parameter of the AI ​​model for each label learned by the AI ​​model using the code corresponding to the label; determining noise to be added to the parameters of the AI ​​model based on a value indicating the importance of the parameters of the AI ​​model for labels to be forgotten and a value indicating the importance of the parameters of the AI ​​model for labels not to be forgotten; updating the parameters of the AI ​​model by adding the determined noise to the parameters of the AI ​​model; A learning method comprising:

8. A step of training an AI model using a dataset in which each piece of data is assigned a code corresponding to the label of the data; calculating a value indicating the importance of a parameter of the AI ​​model for each label learned by the AI ​​model using the code corresponding to the label; determining noise to be added to the parameters of the AI ​​model based on a value indicating the importance of the parameters of the AI ​​model for labels to be forgotten and a value indicating the importance of the parameters of the AI ​​model for labels not to be forgotten; updating the parameters of the AI ​​model by adding the determined noise to the parameters of the AI ​​model; A learning program that allows a computer to execute the above.

Images