Model management device, model management system and model management method

Abstract

Disclosed is technology for managing training data used to train a machine-learning model. One aspect of the present disclosure relates to a model management device comprising: a data identifying unit that acquires data identifying information; a model identifying unit that, with reference to usage information of training data, identified a first machine-learning model which has been trained by using first training data identified by the data identifying information; and a processing unit that deletes the first training data from a first training data set used to train the first machine-learning model, and thereby creates a second training data set and executes processing the first machine-learning model.

Description

[Technical field]

[0001] The present disclosure relates to a model management device, a model management system, and a model management method. [Background technology]

[0002] With the advancement of machine learning technology such as deep learning, machine learning technology is widely used in various technical fields. It is generally known that the performance of machine learning models can change depending on the training data. It is also known that it is difficult to predict such changes. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] JP 2020-102001 A [Patent Document 2] JP 2021-157313 A Summary of the Invention [Problem to be solved by the invention]

[0004] While it is considered important to manage which training data was used to train a machine learning model, data management techniques for managing the association between a machine learning model and the training data used to train the machine learning model have not received much attention in the past.

[0005] In view of the above problems, one objective of the present disclosure is to provide a technique for managing training data used to train a machine learning model. [Means for solving the problem]

[0006] One aspect of the present disclosure relates to a model management device having a data identification unit that acquires data identification information, a model identification unit that refers to training data usage information and identifies a first machine learning model trained using first training data identified by the data identification information, and a processing unit that creates a second training dataset by deleting the first training data from a first training dataset used to train the first machine learning model, and performs processing on the first machine learning model. Effect of the Invention

[0007] In accordance with the present disclosure, techniques are provided for managing training data used to train machine learning models. [Brief description of the drawings]

[0008] [Figure 1] FIG. 1 is a schematic diagram illustrating a model management system according to one embodiment of the present disclosure. [Diagram 2] FIG. 2 is a block diagram illustrating a hardware configuration of a model management device according to an embodiment of the present disclosure. [Diagram 3] FIG. 3 is a block diagram illustrating a functional configuration of a model management device according to an embodiment of the present disclosure. [Figure 4] FIG. 4 is a diagram illustrating detailed information of training data according to an embodiment of the present disclosure. [Diagram 5] FIG. 5 is a diagram illustrating usage information according to one embodiment of the present disclosure. [Figure 6] FIG. 6 is a flowchart illustrating a model management process according to an embodiment of the present disclosure. [Figure 7] FIG. 7 is a schematic diagram illustrating a model management system according to another embodiment of the present disclosure. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0009] Hereinafter, embodiments of the present disclosure will be described with reference to the drawings.

[0010] In the following embodiments, a model management device for managing training data used to train a machine learning model is disclosed.

[0011] In the embodiments described below, for example, a model management device is disclosed that can deal with a case in which a problem is found regarding the quality of some or all of the training data in a training dataset during operation of a machine learning model trained using the training dataset.

[0012] For example, problems may be considered with regard to the training data from the viewpoint of the performance of the machine learning model. As an example, it may be found after training of the machine learning model that the method of acquiring the training data at a certain facility from among the training datasets used to train the machine learning model is inappropriate or that the training data is fabricated. In this case, the training data acquired at the facility is inappropriate and may degrade the performance of the machine learning model. For this reason, it is necessary to avoid using the training data acquired at the facility, and it is desirable to retrain the machine learning model trained with the training dataset including the inappropriate training data.

[0013] As another example, performance may decline because the environment in which the data was acquired differs from the environment in which it is being used. For example, people's lifestyles and behaviors have changed since the COVID-19 pandemic, so training data acquired before the pandemic may not be appropriate for a system that will be used after the pandemic. In this case, the training data acquired before the pandemic is inappropriate and may degrade the performance of the machine learning model. For this reason, the use of such training data acquired before the pandemic should be avoided, and it is desirable to retrain a machine learning model that has been trained using a training dataset that includes inappropriate training data.

[0014] As yet another example, it may be found after training of a machine learning model that annotations (e.g., annotations, labels, etc.) made by a certain vendor are of poor quality. In this case, the training data annotated by the vendor is inappropriate and may degrade the performance of the machine learning model. For this reason, it is necessary to avoid using the training data annotated by the vendor, and it is desirable to retrain the machine learning model trained with a training dataset including inappropriate training data.

[0015] That is, the quality of the training data used to train the machine learning model needs to be judged, and data that does not meet the criteria, is questionable, or differs from the current environment needs to be excluded from training in terms of the machine learning model's performance.

[0016] Alternatively, there may be legal and / or ethical issues with regard to the training data. For example, it may be discovered after training of a commercial machine learning model that training data was used to train the model that was not permitted or licensed for commercial use, or that was previously permitted or licensed but that permission or license has since expired. In this case, from a contractual perspective, the use of the unauthorized training data should be avoided, and it is desirable to retrain the machine learning model that was trained with a training dataset that includes the unauthorized training data.

[0017] Another example may be a case where training data from an individual who has not consented to the use of their data or who previously consented to the use of their data but has since requested to revoke their consent has been used to train a machine learning model, in which case, from the perspective of privacy and / or privacy laws, it may be necessary to avoid using training data that includes personal information from which consent has not been granted, and it may be desirable to retrain a machine learning model that has been trained with a training dataset that includes such training data.

[0018] As another example, when developing artificial intelligence (AI) for recruitment, training data including age and / or gender may be used to train a machine learning model. In this case, from an ethical point of view, the use of training data including age and / or gender should be avoided, and it is desirable to retrain a machine learning model trained with a training dataset including such training data using a dataset that excludes inappropriate data.

[0019] [Summary] In summary, when it is determined that such inappropriate training data has been used to train a machine learning model, a model management device according to one embodiment of the present disclosure may identify the machine learning model that was trained using the inappropriate training data, remove the inappropriate training data from the training dataset, and retrain the identified machine learning model using a training dataset that does not include the inappropriate training data.

[0020] A model management system 10 according to the following embodiment includes a training data database (DB) 20, a usage information database (DB) 30, a terminal 40, and a model management device 100, as shown in FIG.

[0021] The training data DB 20 stores training data. Specifically, the training data DB 20 stores training data used to train one or more machine learning models managed by the model management device 100. For example, the training data DB 20 may store detailed information, such as the acquisition location and acquisition time of each training data, annotation information, license information, personal information protection, and ethics information, in association with the training data.

[0022] The usage information DB30 stores usage information of the training data. Here, the usage information may indicate an association between a machine learning model and the training data used to train the machine learning model. Specifically, the usage information DB30 stores usage information indicating the usage status of the training data for one or more machine learning models managed by the model management device 100. For example, the usage information DB30 may store, in association with each machine learning model managed by the model management device 100, identification information of the training data used to train the machine learning model.

[0023] The terminal 40 may be a personal computer (PC), a tablet, a smartphone, or the like, operated by a user such as an administrator of a machine learning model. The terminal 40 is connected to the model management device 100 by wire or wirelessly, and the user can operate the terminal 40 to execute various processes described later for the model management device 100. The terminal 40 may also receive data identification information, such as identification information for identifying inappropriate data, from the user and provide the data identification information to the model management device 100.

[0024] When the model management device 100 acquires data identification information indicating inappropriate training data or the like from the terminal 40, it refers to the usage information of the training data stored in the usage information DB 30 and identifies a machine learning model trained with the training data identified by the data identification information. Then, the model management device 100 updates the training dataset by deleting the identified training data from the training dataset stored in the training data DB 20 that was used to train the identified machine learning model, performs processing on the identified machine learning model (e.g., retraining the machine learning model with the updated training dataset), and reports the processing result to the terminal 40.

[0025] Here, the model management device 100 may be realized by a computing device such as a server, a personal computer (PC), a smartphone, or a tablet, and may have a hardware configuration such as that shown in Fig. 2. That is, the model management device 100 has 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.

[0026] A program or instructions for implementing various functions and processes described below in the model management device 100 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 program or instructions are installed from the storage medium via the drive device 101 into the storage device 102 or the memory device 103. However, the program or instructions do not necessarily have to be installed from the storage medium, and may be downloaded from any external device via a network or the like.

[0027] The storage device 102 is realized by a hard disk drive or the like, and stores installed programs or instructions as well as files, data, etc. used in executing the programs or instructions.

[0028] The memory device 103 is realized by a random access memory, a static memory, or the like, and when a program or an instruction is activated, reads and stores the program or instruction, data, or the like from the storage device 102. The storage device 102, the memory device 103, and the removable storage medium may be collectively referred to as a non-transitory tangible storage medium.

[0029] The processor 104 may be realized by one or more CPUs (Central Processing Units), GPUs (Graphics Processing Units), processing circuitry, etc., which may be composed of one or more processor cores, and performs various functions and processing of the model management device 100 described below in accordance with programs, instructions, data such as parameters necessary to execute the programs or instructions, etc. stored in the memory device 103.

[0030] The user interface (UI) device 105 may be composed of 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 realizes an interface between a user and the model management device 100. For example, a user operates the model management device 100 by operating a GUI (Graphical User Interface) displayed on a display or touch panel using a keyboard, mouse, etc.

[0031] The communication device 106 is realized by various communication circuits that execute wired and / or wireless communication processing with an external device, the Internet, a LAN (Local Area Network), a cellular network, or other communication network.

[0032] However, the above-described hardware configuration is merely an example, and the model management device 100 according to the present disclosure may be realized by any other appropriate hardware configuration.

[0033] [Model Management Device] Next, a model management device 100 according to an embodiment of the present disclosure will be described with reference to Figures 3 to 5. The model management device 100 according to this embodiment may refer to the usage information of the training data, identify a machine learning model trained with the training data identified by the data identification information, delete the identified training data from a training dataset used to train the identified machine learning model, and execute a process (e.g., retraining) on ​​the machine learning model with an updated training dataset.

[0034] Fig. 3 is a block diagram showing a functional configuration of a model management device 100 according to an embodiment of the present disclosure. As shown in Fig. 3, the model management device 100 includes a data specifying unit 110, a model specifying unit 120, and a processing unit 130. For example, one or more functional units of the data specifying unit 110, the model specifying unit 120, and the processing unit 130 may be realized by one or more processors 104 executing one or more programs or instructions stored in a non-transitory tangible storage medium such as the storage device 102 and / or the memory device 103.

[0035] The data specifying unit 110 acquires data identification information. Specifically, when acquiring identification information for identifying training data from the terminal 40 or the like, the data specifying unit 110 identifies training data corresponding to the identification information from the training data set stored in the training data DB 20 based on the acquired identification information.

[0036] For example, the data identification information may indicate training data that was used to train a machine learning model managed by the model management device 100 but was found to be inappropriate data after training. Specifically, the data identification information may indicate training data that may degrade the performance of the machine learning model by using the training data. For example, such training data may be training data acquired by an inappropriate acquisition method, fabricated training data, training data acquired before the COVID-19 pandemic, training data annotated by a bad company, etc. The data identification information may also indicate training data whose use may be problematic from a legal and / or ethical perspective. For example, such training data may include training data that is not permitted or licensed for commercial use, training data that was previously permitted or licensed but has since been invalidated, training data related to an individual who has not authorized the use of data, training data related to an individual who previously authorized the use of data but has since requested to revoke authorization, and training data including age and / or gender.

[0037] The data identification information may be identification information for identifying individual training data, but is not limited thereto. For example, the data identification information may be provided in the form of a data identification condition (e.g., indicating a specific acquisition location, acquisition time, annotation company, etc.) for identifying or extracting a plurality of training data stored in the training data DB 20. When such a data identification condition is acquired from the terminal 40, the data identification unit 110 may identify a training data set in the training data DB 20 that corresponds to the acquired data identification condition.

[0038] For example, when the training data DB 20 holds detailed information for each training data as shown in Fig. 4, the data identifying unit 110 may search for training data that meets the data identifying condition and extract the detected training data set as the training data that meets the data identifying condition. For example, when the data identifying condition is the acquisition location "facility A", the data identifying unit 110 may extract training data #001, #005, and #008 that meet the acquisition location "facility A". Also, when the data identifying condition is personal information protection "permission", the data identifying unit 110 may extract training data #001, #004, #006, and #008 that meet the personal information protection "permission".

[0039] The model identification unit 120 refers to the usage information of the training data to identify a machine learning model trained using the training data identified by the data identification information. Specifically, when the training data corresponding to the data identification information acquired from the terminal 40 or the like is identified, the model identification unit 120 accesses the usage information stored in the usage information DB 30 and identifies the machine learning model trained using the identified training data.

[0040] For example, the usage information may have a table-format data structure as shown in Fig. 4, and may have, for example, two columns: a "model index" and a "data index." The "model index" identifies the machine learning models #X1, #X2, ... managed by the model management device 100. In addition, the "data index" identifies the training data #001, #002, ... stored in the training data DB 20, and indicates the training data used to train each machine learning model.

[0041] For example, when training data #001 is identified by the data identification information, the model identification unit 120 may refer to the usage information to identify a model index #X1 corresponding to the data index #001. Also, when training data #002 is identified by the data identification information, the model identification unit 120 may refer to the usage information to identify a model index #X2 corresponding to the data index #002.

[0042] The processing unit 130 creates a new training dataset by deleting the training data identified by the data identification information from the training dataset used to train the machine learning model, and performs processing on the machine learning model. Specifically, when a machine learning model trained using the training data identified by the acquired data identification information is identified, the processing unit 130 may delete the identified training data from the training dataset used to train the identified machine learning model, and update the training dataset with the remaining training data. Then, the processing unit 130 may perform processing on the machine learning model based on the updated training dataset, for example, may retrain the machine learning model using the updated training dataset.

[0043] For example, as described above, if the training data #001 is identified as inappropriate data, the processing unit 130 may delete the training data #001 from the training datasets #001, #004, #006, #008, ... used to train the machine learning model #X1, and retrain the machine learning model #1 with the updated training datasets #004, #006, #008, .... Similarly, if the training data #002 is identified as inappropriate data, the processing unit 130 may delete the training data #002 from the training datasets #002, #003, #005, #007, #009, ... used to train the machine learning model #X2, and retrain the machine learning model #2 with the updated training datasets #003, #005, #007, #009, ....

[0044] Furthermore, the processing unit 130 may delete the training data identified by the data identification information from the training data DB 20, or may make the training data unusable. This makes it possible to avoid using the unusable training data, and also to free up the storage area reserved for the unusable training data.

[0045] The processing unit 130 may also determine whether the machine learning model is usable at a given time. Specifically, when data identification information indicating the expiration date of the training data is acquired, such as when the license M expires, the data identification unit 110 refers to detailed information of the training data DB 20 to identify the training data #001, #004, #006, and #008 having the license information "License M". Since the identified training data #001, #004, #006, and #008 become unusable after the expiration date, the processing unit 130 may determine that the machine learning model trained using the training data #001, #004, #006, and #008 becomes unusable after the expiration date of the license M. In this case, the processing unit 130 may retrain the machine learning model with a training dataset from which the training data #001, #004, #006, and #008 have been deleted, and operate the retrained machine learning model after the expiration date of the license M. This enables the operation of machine learning models to be properly managed in accordance with license expiration dates.

[0046] According to this embodiment, a machine learning model trained using training data identified as inappropriate data can be retrained with a training dataset from which the inappropriate data has been deleted, and a machine learning model that is suitable not only from the perspective of the performance of the machine learning model but also from a legal and / or ethical perspective can be reacquired.

[0047] [Model Management Process] Next, a model management process according to an embodiment of the present disclosure will be described with reference to Fig. 6. The model management process is executed by the above-mentioned model management device 100, and more specifically, may be realized by one or more processors 104 of the model management device 100 executing one or more programs or instructions stored in one or more memory devices 103. Fig. 6 is a flowchart showing the model management process according to an embodiment of the present disclosure.

[0048] As shown in Fig. 6, in step S101, the model management device 100 acquires data identification information. Specifically, the model management device 100 may acquire data identification information for identifying one or more inappropriate training data from a user via the terminal 40 or the like. For example, the data identification information may be identification information for identifying training data, or may be a data identification condition for identifying one or more training data. For example, when a data identification condition that "the acquisition location is facility A" is provided, the model management device 100 may refer to detailed information in the training data DB 20 and identify training data #001, #005, and #008 whose "acquisition location" corresponds to "facility A".

[0049] In step S102, the model management device 100 refers to the usage information to identify machine learning models trained using the identified training data. Specifically, the model management device 100 may refer to the usage information stored in the usage information DB 30 to identify machine learning models trained using each training data identified in step S101. For example, the model management device 100 may identify two machine learning models #X1 and #X2 as machine learning models trained using the training data #001, #005, and #008.

[0050] In step S103, the model management device 100 deletes the identified training data from the training dataset used to train the identified machine learning model. For example, when the machine learning models #X1 and #X2 are identified in step S102, the model management device 100 identifies the training datasets #TD1={#001, #004, #006, #008,} and #TD2={#002, #003, #005, #007, #009,} used to train the machine learning models #X1 and #X2, respectively. Then, the model management device 100 deletes the training data #001, #005, and #008 from the training datasets #TD1 and #TD#2, and creates #TD1_deleted={#004, #006,} and #TD2_deleted={#002, #003, #007, #009,}.

[0051] In step S104, the model management device 100 executes processing on the machine learning models. For example, the model management device 100 may retrain the machine learning models #X1 and #X2 using #TD1_deleted={#004, #006,} and #TD2_deleted={#002, #003, #007, #009,}, respectively.

[0052] According to this embodiment, a machine learning model trained using training data identified as inappropriate data can be retrained with a training dataset from which the inappropriate data has been deleted, and a machine learning model that is suitable not only from the perspective of the performance of the machine learning model but also from a legal and / or ethical perspective can be reacquired.

[0053] [Variations] Next, a model management system 10 according to another embodiment of the present disclosure will be described with reference to Fig. 7. The model management system 10 according to this embodiment further includes a data conversion device 50 and a conversion database (DB) 60.

[0054] The data conversion device 50 executes data conversion for deleting or concealing personal information from the training data including personal information stored in the training data DB 20, and stores the converted training data in the conversion DB 60.

[0055] The conversion DB 60 stores the training data converted by the data conversion device 50 so as to conceal or delete personal information from the data. The model management device 100 may train a machine learning model using the converted training data stored in the conversion DB 60 instead of the training data stored in the training data DB 20. This allows the model management device 100 to train a machine learning model using training data that does not include personal information. Note that the conversion DB 60 may store detailed information of each training data stored in the training data DB 20 in association with the converted training data.

[0056] In the model management system 10 according to the present embodiment, when data identification information for identifying inappropriate data is acquired, the data identification unit 110 identifies the inappropriate training data identified by the data identification information, and the model identification unit 120 identifies a machine learning model trained using the identified inappropriate training data by referring to the usage information. Then, the processing unit 130 identifies the converted training dataset used to train the identified machine learning model in the conversion DB 60, and deletes the inappropriate training data from the identified converted training dataset. Furthermore, the processing unit 130 may retrain the machine learning model using the converted training dataset from which the inappropriate training data has been deleted.

[0057] According to this embodiment, data conversion is performed to delete or conceal personal information from training data that includes personal information, and the converted training data is used to train a machine learning model. Even if inappropriate training data is found during operation of the machine learning model, the inappropriate training data can be deleted and the machine learning model can be retrained with updated training data that does not include personal information.

[0058] In addition, the following supplementary notes are provided in relation to the above description. (Appendix 1) A data identification unit that acquires data identification information; a model identification unit that identifies a first machine learning model trained using the first training data identified by the data identification information by referring to the training data usage information; a processor for generating a second training data set by deleting the first training data from a first training data set used to train the first machine learning model, and for performing processing on the first machine learning model; A model management device comprising: (Appendix 2) 2. The model management device of claim 1, wherein the usage information indicates an association between a machine learning model and training data used to train the machine learning model. (Appendix 3) The model management device of claim 1 or 2, wherein the training data is associated with one or more of the location of acquisition, the time of acquisition, annotation information, license information, and personal information protection and ethics information. (Appendix 4) 4. The model management device of claim 1, wherein the processing unit retrains the first machine learning model using the second training data set. (Appendix 5) 5. The model management device of claim 1, wherein the training data is generated by anonymizing or deleting personal information from data. (Appendix 6) 6. The model management device of claim 1, wherein the processing unit determines whether the first machine learning model is usable at a given point in time. (Appendix 7) A training data database (DB) for storing training data; A usage information database (DB) for storing usage information of the training data; a model management device communicatively connected to the training data DB and the usage information DB; having The model management device includes: A data identification unit that acquires data identification information; a model identification unit that identifies a first machine learning model trained using the first training data identified by the data identification information by referring to the usage information; a processor for generating a second training data set by deleting the first training data from a first training data set used to train the first machine learning model, and for performing processing on the first machine learning model; A model management system having: (Appendix 8) Obtaining data specific information; referencing training data usage information to identify a first machine learning model trained using the first training data identified by the data identification information; creating a second training data set by removing the first training data from a first training data set used to train the first machine learning model, and performing processing on the first machine learning model; 1. A computer implemented method for model management comprising:

[0059] 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.

[0060] The disclosures of the specification, drawings and abstract contained in Japanese application No. 2022-093639, filed on June 9, 2022, are incorporated herein by reference in their entirety. [Explanation of symbols]

[0061] 10. Model Management System 20 Training Data DB 30 Usage information DB 40 Terminals 50 Data conversion device 60 Conversion DB 100 Model conversion device 110 Data Identification Section 120 Model Specifications 130 Processing section

Claims

1. A data identification unit that acquires data identification information; a model identification unit that identifies a first machine learning model trained using the first training data identified by the data identification information by referring to the training data usage information; a processor that creates a second training data set by deleting the first training data from a first training data set used to train the first machine learning model, and performs processing on the first machine learning model; A model management device comprising:

2. The model management apparatus of claim 1 , wherein the usage information indicates an association between a machine learning model and training data used to train the machine learning model.

3. The model management device according to claim 1 , wherein the training data is associated with one or more of an acquisition location, an acquisition time, annotation information, license information, and personal information protection and ethics information.

4. The model management apparatus according to claim 1 , wherein the processing unit retrains the first machine learning model using the second training data set.

5. The model management device according to claim 1 , wherein the training data is generated by anonymizing or deleting personal information from the data.

6. The model management apparatus according to claim 1 , wherein the processing unit determines whether the first machine learning model is usable at a given point in time.

7. A training data database (DB) for storing training data; A usage information database (DB) for storing usage information of the training data; a model management device communicatively connected to the training data DB and the usage information DB; having The model management device includes: A data identification unit that acquires data identification information; a model identification unit that identifies a first machine learning model trained using the first training data identified by the data identification information by referring to the usage information; a processor that creates a second training data set by deleting the first training data from a first training data set used to train the first machine learning model, and performs processing on the first machine learning model; A model management system having:

8. Obtaining data specific information; referencing training data usage information to identify a first machine learning model trained using the first training data identified by the data identification information; creating a second training data set by removing the first training data from a first training data set used to train the first machine learning model, and performing processing on the first machine learning model; 1. A computer implemented method for model management comprising:

Images