Training device, training method, machine learning model, authentication device, authentication method and program

Abstract

To provide a method and the like for training a machine learning model to identify an identification object person from image data with excellent accuracy while taking attributes of the identification object person into consideration.SOLUTION: A method acquires a training data set composed of training image data for training a machine learning model composed of a first sub model for generating a feature amount of image data of an identification object person and a second sub model for generating an identification result of the identification object person from the feature amount, and an identification label, inputs the training image data to the machine learning model, executes partition learning for adjusting a parameter of the second sub model according to an error between the identification result from the machine learning model and the identification label, classifies the training image data by partition, inputs the training image data to the machine learning model for each partition, and executes feature learning for adjusting a parameter between the first sub model and the second sub model according to the error between the identification result from the machine learning model and the identification label.SELECTED DRAWING: Figure 6

Description

[Technical field]

[0001] The present disclosure relates to a training device, a training method, a machine learning model, an authentication device, an authentication method, and a program. [Background technology]

[0002] With the recent advances in deep learning technology, machine learning models are being used in a wide range of technical fields. For example, image processing models that can be realized using convolutional neural networks are widely used in object detection, anomaly detection, image classification, biometric authentication, and more.

[0003] For example, in face recognition, a machine learning model is trained using a huge amount of face image data. The face image data used is collected in accordance with the world population ratio of each race, and typically, the proportion of face image data of Caucasians is relatively high, while the proportion of face image data of Indians and Africans may be relatively low. Similarly, the proportion of face image data of men may be relatively high, while the proportion of face image data of women may be relatively low.

[0004] In a typical machine learning model, there may be a difference in accuracy of face recognition between attributes with a lot of training data (e.g., Caucasian and male) and attributes with little training data (e.g., Indian or African and female). In other words, face image data with more training data attributes may be recognized with higher accuracy, and face image data with fewer training data attributes may be recognized with lower accuracy. Such demographic attributes of the training data may affect the accuracy of the face recognition model. [Prior art documents] [Non-patent literature]

[0005] [Non-Patent Document 1] Invariant Risk Minimization (https: / / arxiv.org / abs / 1907.02893) Summary of the Invention [Problem to be solved by the invention]

[0006] However, adding such demographic attribute labels to training data can significantly increase the cost of creating the training data. For this reason, an approach is desired for training a machine learning model that can identify targets with high accuracy from training data consisting of image data and target target identification labels without adding attribute labels.

[0007] In view of the above problems, one objective of the present disclosure is to provide a technique for training a machine learning model for identifying a subject to be identified from image data with high accuracy while taking into account the attributes of the subject to be identified. [Means for solving the problem]

[0008] One aspect of the present disclosure relates to a training device having: a data acquisition unit that acquires a training dataset consisting of training image data and an identification label for training a machine learning model consisting of a first sub-model that generates features from image data of a person to be identified and a second sub-model that generates an identification result of the person to be identified from the features; a partition learning unit that inputs the training image data to the machine learning model and performs partition learning to adjust parameters of the second sub-model according to an error between the identification result from the machine learning model and the identification label; and a feature learning unit that classifies training image data of the training dataset by partition, inputs the training image data by partition to the partition-trained machine learning model, and performs feature learning to adjust parameters of the first sub-model and the second sub-model according to an error between the identification result from the partition-trained machine learning model and the identification label. Effect of the Invention

[0009] According to the present disclosure, it is possible to train a machine learning model for identifying a target person from image data with good accuracy while taking into account the attributes of the target person. [Brief description of the drawings]

[0010] [Figure 1] 1A and 1B are schematic diagrams illustrating the training and inference processes of a face recognition model according to one embodiment of the present disclosure. [Diagram 2] FIG. 2 is a schematic diagram illustrating a training device according to one embodiment of the present disclosure. [Diagram 3] FIG. 3 is a block diagram illustrating a hardware configuration of a training device and an authentication device according to an embodiment of the present disclosure. [Figure 4] FIG. 4 is a schematic diagram illustrating partition learning and feature learning according to one embodiment of the present disclosure. [Diagram 5] FIG. 5 is a block diagram illustrating a functional configuration of a training device according to an embodiment of the present disclosure. [Figure 6] FIG. 6 is a schematic diagram illustrating a backbone sub-model and a classification sub-model according to one embodiment of the present disclosure. [Figure 7] FIG. 7 illustrates pseudocode for a training algorithm according to one embodiment of the present disclosure. [Figure 8] FIG. 8 is a flowchart illustrating partition learning according to one embodiment of the present disclosure. [Figure 9] FIG. 9 is a flowchart illustrating feature learning according to one embodiment of the present disclosure. [Figure 10] FIG. 10 is a schematic diagram illustrating an authentication device according to an embodiment of the present disclosure. [Figure 11] FIG. 11 is a block diagram illustrating a functional configuration of an authentication device according to an embodiment of the present disclosure. [Figure 12] FIG. 12 is a flowchart illustrating an authentication process according to an embodiment of the present disclosure. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0011] Hereinafter, embodiments of the present disclosure will be described with reference to the drawings.

[0012] In the following embodiments, a training device that trains a machine learning model to identify a subject to be identified in image data and an authentication device that authenticates a subject to be authenticated using the trained machine learning model are disclosed.

[0013] [Facial recognition model] 1A, the face recognition model 20 to be trained is composed of a backbone sub-model (BACKBONE) and a classification sub-model (CLASSIFIER), and when it receives face image data for training, it outputs a recognition result of the person to be identified in the face image data. Here, the backbone sub-model may generate features from the face image data of the person to be identified, and the classification sub-model may generate a recognition result of the person to be identified from the features generated by the backbone sub-model.

[0014] On the other hand, as shown in Fig. 1B, the trained face recognition model 30 is composed of a backbone sub-model obtained by training the face recognition model 20 in a training process, and upon receiving face image data of a target for authentication, outputs the feature amount of a person to be authenticated in the face image data. For example, if the output feature amount is identical to or similar to the feature amount of a registered person, the person to be authenticated may be determined to be the registered person. On the other hand, if the output feature amount is dissimilar to any of the registered people, the person to be authenticated may be determined to be unregistered.

[0015] The face recognition model 20 to be trained may be trained by a training device 100 as shown in FIG. 2. Specifically, the training device 100 acquires a training dataset of training data consisting of face image data for training and an identification label from a training data database (DB) 10, and inputs the face image data to the face recognition model 20 to be trained. Then, when an identification result is acquired from the face recognition model 20, the training device 100 executes partition learning as described in detail below, and adjusts parameters of the classification submodel according to an error between the acquired identification result and the identification label without updating parameters of the backbone submodel. For example, the training device 100 may end the partition learning when it adjusts parameters of the classification submodel for all training data of the training dataset. According to the face recognition model 20 including the classification submodel trained by such partition learning, the face image data for training may be automatically classified according to, for example, demographic attributes.

[0016] Thereafter, the training device 100 executes feature learning as described in detail below, classifies the training face image data for each partition, and inputs the face image data for each partition to the face recognition model 20 including the classification sub-model trained by the partition learning. For example, the training device 100 may first input face image data classified as partition #1 (e.g., Caucasian, etc.) to the face recognition model 20, and then input face image data classified as partition #2 (e.g., Indian, etc.) to the face recognition model 20. Then, when the identification result is obtained from the face recognition model 20, the training device 100 adjusts parameters of the backbone sub-model and the classification sub-model according to the error between the obtained identification result and the identification label. For example, the training device 100 may end the feature learning when it adjusts parameters of the backbone sub-model and the classification sub-model for all training data of the training dataset. According to the backbone sub-model trained by such feature learning, the training face image data can be converted into a feature amount that takes into account, for example, attributes, etc. Therefore, the authentication device 200 using the trained backbone submodel may be able to achieve face authentication accuracy for attributes with a small amount of training data similar to that for attributes with a large amount of training data.

[0017] Here, the training device 100 and the authentication device 200 are realized by a computing device such as a server, and may have a hardware configuration as shown in Fig. 3. That is, the training device 100 and the authentication device 200 have a drive device 101, a storage device 102, a memory device 103, a processor 104, a user interface (UI) device 105, and a communication device 106, which are interconnected via a bus B.

[0018] The programs or instructions for implementing various functions and processes described later in the training device 100 and the authentication device 200 may be stored in a removable storage medium such as a CD-ROM (Compact Disk-Read Only Memory) or a flash memory. When the storage medium is set in the drive device 101, the programs or instructions are installed from the storage medium to the storage device 102 or the memory device 103 via the drive device 101. However, the programs or instructions do not necessarily need to be installed from the storage medium, and may be downloaded from any external device via a network or the like. The storage device 102 is realized by a hard disk drive or the like, and stores files, data, etc. used to execute the programs or instructions together with the installed programs or instructions. The memory device 103 is realized by a random access memory, a static memory, etc., and when the programs or instructions are activated, the programs, instructions, data, etc. are read from the storage device 102 and stored. The storage device 102, the memory device 103, and the removable storage medium may be collectively referred to as a non-transitory storage medium.

[0019] The processor 104 may be realized by one or more central processing units (CPUs), graphics processing units (GPUs), processing circuitry, etc., each of which may be configured with one or more processor cores, and executes various functions and processes of the training device 100 and the authentication device 200 described below according to programs, instructions, data such as parameters necessary to execute the programs or instructions, etc., stored in the memory device 103. The user interface (UI) device 105 may be configured by input devices such as a keyboard, mouse, camera, microphone, etc., output devices such as a display, speaker, headset, printer, etc., and input / output devices such as a touch panel, and implements an interface between the user and the training device 100 and the authentication device 200. For example, the user operates the training device 100 and the authentication device 200 by operating a graphical user interface (GUI) displayed on a display or a touch panel using a keyboard, mouse, etc. The communication device 106 is realized by various communication circuits that execute communication processes with external devices, the Internet, a communication network such as a LAN (Local Area Network), etc.

[0020] However, the above-mentioned hardware configuration is merely an example, and the training device 100 and the authentication device 200 according to the present disclosure may be realized by any other suitable hardware configuration.

[0021] [Training device] Next, a training device 100 according to an embodiment of the present disclosure will be described. In this embodiment, the training device 100 performs partition learning to divide the training data of a training data set according to attributes, and then performs feature learning using the training data for each partition. Note that the following embodiment focuses on a face recognition model 20 realized by a neural network as a machine learning model to be trained, but the machine learning model according to the present disclosure is not necessarily limited to the face recognition model 20, and may be, for example, any other type of machine learning model that identifies an identification target from image data of various attributes.

[0022] For example, in the embodiment shown in Fig. 4, training data is first classified into Caucasian and non-Caucasian by partition learning (Partition Learning 1). Then, face recognition model 20 is trained using Caucasian training data and non-Caucasian training data by feature learning (Feature Learning 1). Face recognition model 20 thus trained for each race can realize face recognition with similar accuracy for each race, regardless of the amount of training data for each race.

[0023] Furthermore, the training data is classified into male and female by partition learning (Partition Learning 2), and the facial recognition model 20 is further trained by using the male training data and the female training data by feature learning (Feature Learning 2). The facial recognition model 20 that has been feature-learned for each gender in this way can achieve facial recognition with similar accuracy not only for race but also for gender, regardless of the amount of training data for each gender.

[0024] Fig. 5 is a block diagram showing a functional configuration of a training device 100 according to an embodiment of the present disclosure. As shown in Fig. 5, the training device 100 includes a data acquisition unit 110, a partition learning unit 120, and a feature learning unit 130. For example, one or more functional units of the data acquisition unit 110, the partition learning unit 120, and the feature learning unit 130 may be realized by one or more processors 104 executing one or more programs or instructions.

[0025] The data acquisition unit 110 acquires a training data set consisting of training image data and identification labels for training a machine learning model consisting of a first sub-model that generates a feature amount from image data of a person to be identified and a second sub-model that generates an identification result of the person to be identified from the feature amount. The data acquisition unit 110 then provides the acquired training data set to the partition learning unit 120 and the feature learning unit 130.

[0026] For example, the machine learning model to be trained may be composed of a backbone sub-model that generates features from facial image data of a person to be identified, and a classification sub-model that determines the identification result of the person to be identified (e.g., identification (ID) information of the person to be identified, etc.) from the features generated by the backbone sub-model. Also, the training dataset may include training data composed of a pair of facial image data of the person to be identified and an identification label of the person to be identified (e.g., identification (ID) information of the person to be identified, etc.).

[0027] The partition learning unit 120 executes partition learning in which training image data is input to the machine learning model, and parameters of the second sub-model are adjusted according to an error between a classification result from the machine learning model and an identification label (step 1). Specifically, the partition learning unit 120 inputs training face image data to the face recognition model 20 to be trained, and acquires a classification result from the face recognition model 20. When adjusting the parameters of the face recognition model 20 according to an error between the acquired classification result and an identification label, the partition learning unit 120 adjusts only the parameters of the classification sub-model according to the error, without adjusting the parameters of the backbone sub-model of the face recognition model 20.

[0028] For example, the partition learning unit 120 may fix the parameters of the backbone submodel while learning the following loss L inv The partition matrix P that maximizes * Request.

number

[0029] L inv When calculating (Φ, f, P), we use the supervised contrastive loss (SCL) L cls The feature quantity in each subset of the partition matrix P may be evaluated according to the following formula (2):

number

number

[0030] According to formula (2), the partition learning unit 120 may adjust the parameters of the classification sub-model so that the classification results between the training face image data and the face image data of the same identification target as the training face image data are similar, and the classification results between the training face image data and other training face image data of an identification target different from the training face image data are dissimilar. Here, the face image data of the same identification target as the training face image data may be generated by, for example, performing data extension on the training face image data, and specifically, may be rotated, enlarged, reduced, inverted, or the like. Note that the above-mentioned SCL is merely one example of a loss function, and other loss functions may be used.

[0031] In partition learning, the parameters of the backbone sub-model of the face recognition model 20 shown in step 1 of Fig. 6 remain fixed. Note that initially, the parameters of the backbone sub-model and the classification sub-model of the face recognition model 20 may be set randomly. As a result, in partition learning, the face recognition model 20 can be trained by focusing on the classification model for classifying face image data based on attributes of the face image data, etc.

[0032] When a predetermined termination condition for partition learning is satisfied, the partition learning unit 120 terminates partition learning for the face recognition model 20, i.e., parameter update processing for the classification submodel, and provides the face recognition model 20 consisting of the classification submodel with updated parameters and the backbone submodel with fixed parameters to the feature learning unit 130. Here, the predetermined termination condition may be, for example, that the error has converged within a predetermined threshold range, that all prepared training data has been processed, etc.

[0033] The feature learning unit 130 classifies the training image data of the training data set for each partition, inputs the training image data for each partition to the partition-trained machine learning model, and performs feature learning to adjust parameters of the first sub-model and the second sub-model according to an error between a classification result from the partition-trained machine learning model and an identification label (step 2). Specifically, the feature learning unit 130 inputs the training face image data to the face recognition model 20 trained by the partition learning unit 120, and acquires a classification result from the face recognition model 20. The feature learning unit 130 may adjust parameters of the backbone sub-model and the classification sub-model according to an error between the acquired classification result and an identification label.

[0034] For example, the feature learning unit 130 may classify training face image data for each partition using the face recognition model 20 composed of a classification submodel with updated parameters and a backbone submodel with fixed parameters, and input the face image data classified for each partition to the face recognition model 20. For example, the feature learning unit 130 may input each training face image data classified into partition #1 to the face recognition model 20, and adjust the parameters of the backbone submodel and the classification submodel according to the error between the identification result from the face recognition model 20 and the identification label. When the parameter update process using the training face image data classified into partition #1 is completed, the feature learning unit 130 may input each training face image data classified into partition #2 to the face recognition model 20, and adjust the parameters of the backbone submodel and the classification submodel according to the error between the identification result from the face recognition model 20 and the identification label. In the same manner, the feature learning unit 130 adjusts the parameters of the backbone submodel and the classification submodel for each partition.

[0035] Specifically, the feature learning unit 130 may optimize the backbone model and the classification model according to equation (3).

number

[0036] When a predetermined termination condition for feature learning is satisfied, the feature learning unit 130 terminates feature learning for the face recognition model 20, i.e., the parameter update process for the backbone model and the classification sub-model, and provides the face recognition model 20 consisting of the backbone model and the classification sub-model with updated parameters as a trained face recognition model to the authentication device 200. Here, the predetermined termination condition may be, for example, that the error has converged within a predetermined threshold range, that all prepared training data has been processed, etc.

[0037] [Training process] The training process according to one embodiment of the present disclosure will now be described: Figure 7 is a pseudocode showing a training algorithm according to one embodiment of the present disclosure.

[0038] As shown in Figure 7, the training face image data {x i ,y i}, Given a randomly initialized backbone submodel Φ, a classification submodel f, and an empty set P (#1), the training algorithm (INV-REG Training) outputs a trained backbone submodel Φ (#2).

[0039] Initially, the training algorithm performs a partition-free training face image data set {x i ,y i} to train the backbone submodel Φ and the classification submodel f (#3).

[0040] Then, the training algorithm repeats the partition learning in step 1 and the feature learning in step 2 until the error converges (#4).

[0041] First, in partition learning (#5), L is calculated for any k (1≦k≦K) according to the above formula (2). cls(Φ,·,P,k) is calculated (#6). Then, while keeping the parameters of the backbone submodel Φ fixed, the partition matrix P * is learned (#7). Furthermore, P is added to the set of partition matrices P. * is added (#8).

[0042] Next, in feature learning (#9), for any k (1≦k≦K), L cls (Φ,f,P,k) is calculated (#10). Then, with the set of partitions P fixed, the backbone submodel Φ and the classification submodel f are trained according to the above formula (3).

[0043] 8 is a flow chart illustrating partition learning according to one embodiment of the present disclosure. The partition learning may be performed by the training device 100 described above, and more particularly, may be realized by one or more processors 104 of the training device 100 executing one or more programs or instructions stored in one or more memory devices 103.

[0044] 8, in step S101, the training apparatus 100 fixes the backbone sub-model in the face recognition model 20 to be trained. That is, the training apparatus 100 does not update the parameters of the backbone sub-model in partition learning.

[0045] In step S102, the training device 100 generates features from the training face image data. Specifically, the training device 100 inputs the training face image data to a backbone sub-model in the face recognition model 20 to be trained, and acquires features from the backbone sub-model.

[0046] In step S103, the training apparatus 100 calculates the SCL according to equation (2), and adjusts the parameters of the classification sub-model based on the calculated SCL.

[0047] In step S104, the training device 100 calculates a partition matrix P that maximizes the calculated SCL according to the formula (1). * Determine the partition matrix P * It can be indicated which partition each piece of identification information belongs to by.

[0048] 9 is a flow chart illustrating feature learning according to one embodiment of the present disclosure. The feature learning may be performed by the training device 100 described above, and more particularly, may be realized by one or more processors 104 of the training device 100 executing one or more programs or instructions stored in one or more memory devices 103.

[0049] 9, in step S201, the training apparatus 100 fixes the partitions, that is, the training apparatus 100 fixes the set P of partition matrices determined in the partition learning.

[0050] In step S202, the training device 100 inputs training face image data for each partition to the face recognition model 20 to be trained, and calculates the error between the classification result output from the face recognition model 20 and the classification label.

[0051] In step S203, the training device 100 calculates the classification loss L for the training face image data for each partition. cls Calculate (Φ, f, P, k).

[0052] In step S204, the training apparatus 100 calculates an Invariant Risk Minimization (IRM) loss L according to the calculated classification loss for each partition according to equation (4). inv Calculate (Φ,f,P).

number

[0053] In step S205, the training apparatus 100 calculates the classification loss Lcls (Φ,f,P,k) and IRM loss L inv (Φ, f, P). Then, the training device 100 updates the parameters of the face recognition model 20 according to the backpropagation algorithm based on the sum of these losses.

[0054] Once the training of the face recognition model 20 is completed in this manner, the training device 100 may provide the trained backbone sub-model to the authentication device 200 as the trained face recognition model 30. That is, the classification sub-model may be used only in the training process and not in the inference process.

[0055] According to the above-mentioned training device 100 and training process, the attributes of the identification target person in the image data are automatically classified, and the classified image data is used to train the face recognition model 20 for each partition. This makes it possible to generate a face recognition model 20 with good accuracy even for attributes for which there is only a small amount of training data.

[0056] [Authentication device] Next, an authentication device 200 according to an embodiment of the present disclosure will be described. As shown in Fig. 10, the authentication device 200 acquires features from face image data of a person to be authenticated using the trained face recognition model 30 provided by the training device 100 described above, and authenticates the person to be authenticated based on whether the acquired features are identical or similar to features of a registered person. As described above, the trained face recognition model 30 may be composed of a backbone submodel trained by the training device 100.

[0057] 11 is a block diagram showing a functional configuration of an authentication device 200 according to an embodiment of the present disclosure. As shown in FIG. 11, the authentication device 200 includes a data acquisition unit 210 and an authentication unit 220.

[0058] The data acquisition unit 210 acquires image data of the person to be authenticated. Specifically, the data acquisition unit 210 acquires facial image data obtained by capturing an image of the face of the person to be authenticated from a camera that is communicatively connected to the authentication device 200, and passes the acquired facial image data to the authentication unit 220.

[0059] The authentication unit 220 uses the authentication model to authenticate the person to be authenticated based on the image data. Specifically, the authentication unit 220 inputs the face image data provided from the data acquisition unit 210 to the trained face authentication model 30, and acquires the feature amount of the captured person to be authenticated from the trained face authentication model 30. For example, in entrance / exit management using face authentication technology, the authentication unit 220 compares the feature amount of the registered person with the acquired feature amount, and determines whether the person to be authenticated captured in the face image data has been registered. When it is determined that the person to be authenticated captured in the face image data has been registered, the authentication unit 220 may, for example, permit the person to enter the room, and when it is determined that the person to be authenticated captured in the face image data has not been registered, the authentication unit 220 may, for example, not permit the person to enter the room.

[0060] [Authentication process] 12 is a flowchart illustrating an authentication process according to an embodiment of the present disclosure. The authentication process may be performed by the above-described authentication device 200, and more specifically, may be realized by one or more processors 104 of the authentication device 200 executing one or more programs or instructions stored in one or more memory devices 103.

[0061] As shown in FIG. 12, in step S301, the authentication device 200 acquires face image data of the person to be authenticated.

[0062] In step S302, the authentication device 200 authenticates the person to be authenticated based on the facial image data by using the authentication model 30. Specifically, the authentication device 200 inputs the facial image data into the authentication model 30 and acquires the feature amount of the person to be authenticated from the authentication model 30. The authentication device 200 determines whether the acquired feature amount is identical or similar to the feature amount of a registered person to be authenticated, and if a person to be authenticated with the same or similar feature amount as the acquired feature amount is registered, the authentication device 200 authenticates the person to be authenticated as a registered person to be authenticated. On the other hand, if a person to be authenticated with the same or similar feature amount as the acquired feature amount is not registered, the authentication device 200 may determine the person to be authenticated as an unregistered person to be authenticated and not authenticate the person.

[0063] According to the above-described authentication device 200 and authentication process, by utilizing the face recognition model 30 trained for each partition corresponding to an attribute, it is possible to perform authentication with good accuracy even for attributes for which there is only a small amount of training data.

[0064] (Appendix 1) a data acquisition unit that acquires a training data set composed of training image data and identification labels for training a machine learning model composed of a first sub-model that generates a feature amount from image data of a person to be identified and a second sub-model that generates an identification result of the person to be identified from the feature amount; A partition learning unit that inputs the training image data to the machine learning model and performs partition learning to adjust parameters of the second sub-model according to an error between a classification result from the machine learning model and a classification label; a feature learning unit that performs feature learning by classifying training image data of the training data set by partition, inputting the training image data by partition to the partition-trained machine learning model, and adjusting parameters of the first sub-model and the second sub-model according to an error between a classification result from the partition-trained machine learning model and a classification label; A training device having the above features. (Appendix 2) The training device described in Appendix 1, wherein the partition learning unit adjusts parameters of the second sub-model so that classification results between first training image data and second training image data of the same subject to be classified as the first training image data are similar, and classification results between the first training image data and third training image data of a subject to be classified different from the first training image data are dissimilar. (Appendix 3) 3. The training apparatus of claim 2, wherein the second training image data is generated by performing data augmentation on the first training image data. (Appendix 4) 4. The training device according to any one of appendices 1 to 3, wherein the partition learning unit classifies training image data of the training data set in accordance with an attribute of the person to be identified. (Appendix 5) The training device according to any one of appendices 1 to 4, wherein the feature learning unit adjusts parameters of the first sub-model and the second sub-model based on a classification loss and an IRM (Invariant Risk Minimization) loss. (Appendix 6) Obtaining a training dataset composed of training image data and identification labels for training a machine learning model composed of a first sub-model that generates a feature amount from image data of a person to be identified and a second sub-model that generates an identification result of the person to be identified from the feature amount; inputting the training image data into the machine learning model, and performing partition learning to adjust parameters of the second sub-model according to an error between a classification result from the machine learning model and a classification label; Classifying training image data of the training dataset by partition, inputting the training image data by partition to the partition-trained machine learning model, and performing feature learning to adjust parameters of the first sub-model and the second sub-model according to an error between a classification result from the partition-trained machine learning model and a classification label; A training method performed by a computer. (Appendix 7) Obtaining a training dataset composed of training image data and identification labels for training a machine learning model composed of a first sub-model that generates a feature amount from image data of a person to be identified and a second sub-model that generates an identification result of the person to be identified from the feature amount; inputting the training image data into the machine learning model, and performing partition learning to adjust parameters of the second sub-model according to an error between a classification result from the machine learning model and a classification label; Classifying training image data of the training dataset by partition, inputting the training image data by partition to the partition-trained machine learning model, and performing feature learning to adjust parameters of the first sub-model and the second sub-model according to an error between a classification result from the partition-trained machine learning model and a classification label; A program that causes a computer to execute the following. (Appendix 8) Obtaining a training dataset composed of training image data and identification labels for training a machine learning model composed of a first sub-model that generates a feature amount from image data of a person to be identified and a second sub-model that generates an identification result of the person to be identified from the feature amount; inputting the training image data into the machine learning model, and performing partition learning to adjust parameters of the second sub-model according to an error between a classification result from the machine learning model and a classification label; Classifying training image data of the training dataset by partition, inputting the training image data by partition to the partition-trained machine learning model, and performing feature learning to adjust parameters of the first sub-model and the second sub-model according to an error between a classification result from the partition-trained machine learning model and a classification label; A machine learning model, generated by running (Appendix 9) A data acquisition unit that acquires image data of a person to be authenticated; an authentication unit that uses an authentication model to authenticate the person to be authenticated based on the image data; having the authentication model is composed of a first sub-model extracted from a machine learning model including a first sub-model that generates a feature amount from image data of a person to be identified, and a second sub-model that generates an identification result of the person to be identified from the feature amount; The machine learning model is Obtaining a training dataset consisting of training image data and identification labels for training the machine learning model; inputting the training image data into the machine learning model, and performing partition learning to adjust parameters of the second sub-model according to an error between a classification result from the machine learning model and a classification label; Classifying training image data of the training dataset by partition, inputting the training image data by partition to the partition-trained machine learning model, and performing feature learning to adjust parameters of the first sub-model and the second sub-model according to an error between a classification result from the partition-trained machine learning model and a classification label; An authentication device generated by executing (Appendix 10) acquiring image data of a person to be authenticated; authenticating the subject based on the image data utilizing an authentication model; The computer executes the authentication model is composed of a first sub-model extracted from a machine learning model including a first sub-model that generates a feature amount from image data of a person to be identified, and a second sub-model that generates an identification result of the person to be identified from the feature amount; The machine learning model is Obtaining a training dataset consisting of training image data and identification labels for training the machine learning model; inputting the training image data into the machine learning model, and performing partition learning to adjust parameters of the second sub-model according to an error between a classification result from the machine learning model and a classification label; Classifying training image data of the training dataset by partition, inputting the training image data by partition to the partition-trained machine learning model, and performing feature learning to adjust parameters of the first sub-model and the second sub-model according to an error between a classification result from the partition-trained machine learning model and a classification label; The authentication method, which is generated by running (Appendix 11) acquiring image data of a person to be authenticated; authenticating the subject based on the image data utilizing an authentication model; on the computer, the authentication model is composed of a first sub-model extracted from a machine learning model including a first sub-model that generates a feature amount from image data of a person to be identified, and a second sub-model that generates an identification result of the person to be identified from the feature amount; The machine learning model is Obtaining a training dataset consisting of training image data and identification labels for training the machine learning model; inputting the training image data into the machine learning model, and performing partition learning to adjust parameters of the second sub-model according to an error between a classification result from the machine learning model and a classification label; Classifying training image data of the training dataset by partition, inputting the training image data by partition to the partition-trained machine learning model, and performing feature learning to adjust parameters of the first sub-model and the second sub-model according to an error between a classification result from the partition-trained machine learning model and a classification label; A program generated by running

[0065] Although the examples of the present disclosure have been described in detail above, the present disclosure is not limited to the specific embodiments described above, and various modifications and variations are possible within the scope of the gist of the present disclosure described in the claims. [Industrial Applicability]

[0066] INDUSTRIAL APPLICABILITY The present disclosure is useful for a training device and method for training a machine learning model to achieve good authentication processing for authentication subjects with various attributes. [Explanation of symbols]

[0067] 10 Training Data DB 20, 30 Face recognition model 100 training equipment 110 Data Acquisition Section 120 Partition Learning Section 130 Feature Learning Unit 200 Authentication Device 210 Data Acquisition Section 220 Authentication Department

Claims

1. a data acquisition unit that acquires a training data set composed of training image data and identification labels for training a machine learning model composed of a first sub-model that generates a feature amount from image data of a person to be identified and a second sub-model that generates an identification result of the person to be identified from the feature amount; A partition learning unit that inputs the training image data to the machine learning model and performs partition learning to adjust parameters of the second sub-model according to an error between a classification result from the machine learning model and a classification label; a feature learning unit that performs feature learning by classifying training image data of the training data set by partition, inputting the training image data by partition to the partition-trained machine learning model, and adjusting parameters of the first sub-model and the second sub-model according to an error between a classification result from the partition-trained machine learning model and a classification label; A training device having the above features.

2. 2. The training device according to claim 1, wherein the partition learning unit adjusts parameters of the second sub-model so that classification results between first training image data and second training image data of the same subject as the first training image data are similar, and classification results between the first training image data and third training image data of a subject different from the first training image data are dissimilar.

3. The training device of claim 2 , wherein the second training image data is generated by performing data augmentation on the first training image data.

4. The training device according to claim 1 , wherein the partition learning unit classifies the training image data of the training data set in accordance with an attribute of the person to be identified.

5. The training apparatus of claim 1 , wherein the feature learning unit adjusts parameters of the first sub-model and the second sub-model based on a classification loss and an invariant risk minimization (IRM) loss.

6. Obtaining a training dataset composed of training image data and identification labels for training a machine learning model composed of a first sub-model that generates a feature amount from image data of a person to be identified and a second sub-model that generates an identification result of the person to be identified from the feature amount; inputting the training image data into the machine learning model, and performing partition learning to adjust parameters of the second sub-model according to an error between a classification result from the machine learning model and a classification label; Classifying training image data of the training dataset by partition, inputting the training image data by partition to the partition-trained machine learning model, and performing feature learning to adjust parameters of the first sub-model and the second sub-model according to an error between a classification result from the partition-trained machine learning model and a classification label; A training method performed by a computer.

7. Obtaining a training dataset composed of training image data and identification labels for training a machine learning model composed of a first sub-model that generates a feature amount from image data of a person to be identified and a second sub-model that generates an identification result of the person to be identified from the feature amount; inputting the training image data into the machine learning model, and performing partition learning to adjust parameters of the second sub-model according to an error between a classification result from the machine learning model and a classification label; Classifying training image data of the training dataset by partition, inputting the training image data by partition to the partition-trained machine learning model, and performing feature learning to adjust parameters of the first sub-model and the second sub-model according to an error between a classification result from the partition-trained machine learning model and a classification label; A program that causes a computer to execute the following.

8. Obtaining a training dataset composed of training image data and identification labels for training a machine learning model composed of a first sub-model that generates a feature amount from image data of a person to be identified and a second sub-model that generates an identification result of the person to be identified from the feature amount; inputting the training image data into the machine learning model, and performing partition learning to adjust parameters of the second sub-model according to an error between a classification result from the machine learning model and a classification label; Classifying training image data of the training dataset by partition, inputting the training image data by partition to the partition-trained machine learning model, and performing feature learning to adjust parameters of the first sub-model and the second sub-model according to an error between a classification result from the partition-trained machine learning model and a classification label; A machine learning model, generated by running

9. A data acquisition unit that acquires image data of a person to be authenticated; an authentication unit that uses an authentication model to authenticate the person to be authenticated based on the image data; having the authentication model is composed of a first sub-model extracted from a machine learning model including a first sub-model that generates a feature amount from image data of a person to be identified, and a second sub-model that generates an identification result of the person to be identified from the feature amount; The machine learning model is Obtaining a training dataset consisting of training image data and identification labels for training the machine learning model; inputting the training image data into the machine learning model, and performing partition learning to adjust parameters of the second sub-model according to an error between a classification result from the machine learning model and a classification label; Classifying training image data of the training dataset by partition, inputting the training image data by partition to the partition-trained machine learning model, and performing feature learning to adjust parameters of the first sub-model and the second sub-model according to an error between a classification result from the partition-trained machine learning model and a classification label; An authentication device generated by executing

10. acquiring image data of a person to be authenticated; authenticating the subject based on the image data utilizing an authentication model; The computer executes the authentication model is composed of a first sub-model extracted from a machine learning model including a first sub-model that generates a feature amount from image data of a person to be identified, and a second sub-model that generates an identification result of the person to be identified from the feature amount; The machine learning model is Obtaining a training dataset consisting of training image data and identification labels for training the machine learning model; inputting the training image data into the machine learning model, and performing partition learning to adjust parameters of the second sub-model according to an error between a classification result from the machine learning model and a classification label; Classifying training image data of the training dataset by partition, inputting the training image data by partition to the partition-trained machine learning model, and performing feature learning to adjust parameters of the first sub-model and the second sub-model according to an error between a classification result from the partition-trained machine learning model and a classification label; The authentication method, which is generated by running

11. acquiring image data of a person to be authenticated; authenticating the subject based on the image data utilizing an authentication model; on the computer, the authentication model is composed of a first sub-model extracted from a machine learning model including a first sub-model that generates a feature amount from image data of a person to be identified, and a second sub-model that generates an identification result of the person to be identified from the feature amount; The machine learning model is Obtaining a training dataset consisting of training image data and identification labels for training the machine learning model; inputting the training image data into the machine learning model, and performing partition learning to adjust parameters of the second sub-model according to an error between a classification result from the machine learning model and a classification label; Classifying training image data of the training dataset by partition, inputting the training image data by partition to the partition-trained machine learning model, and performing feature learning to adjust parameters of the first sub-model and the second sub-model according to an error between a classification result from the partition-trained machine learning model and a classification label; A program generated by running

Images