Re-learning data management system and re-learning data management method
The relearning data management system addresses the lack of data suitability verification in machine learning models by using a distributed ledger with smart contracts to ensure appropriate data acquisition and verification, enhancing model accuracy and reliability.
Patent Information
- Application Number
- JP2021053536
- Authority / Receiving Office
- JP · JP
- Patent Type
- Patents
- Current Assignee / Owner
- Filing Date
- 2021-03-26
- Publication Date
- 2024-12-06
- Estimated Expiration
- 2041-03-26
AI Technical Summary
Existing relearning processes for machine learning models lack verification of data suitability, leading to potential inappropriate outputs, as the appropriateness of acquired data for relearning is unclear and validation is not always valid.
A relearning data management system and method that utilizes a first AI system to obtain retraining data from a verifier computer, storing it in a distributed ledger system with smart contracts to ensure data integrity and facilitate verification of the machine learning model's appropriateness.
Enables post-relearning verification of the machine learning model's suitability, enhancing data fairness and allowing for timely detection and prevention of abnormalities, thereby improving the accuracy and reliability of the model.
Smart Images

Figure 0007595278000001 
Figure 0007595278000002 
Figure 0007595278000003
Abstract
Description
[Technical field]
[0001] The present disclosure relates to a re-learning data management system and a re-learning data management method. [Background technology]
[0002] Japanese Patent Publication No. 2020-86778 (hereinafter referred to as Patent Document 1) discloses learning a new input / output relationship for a machine-learned learning model and updating the learning model. [Prior art documents] [Patent documents]
[0003] [Patent Document 1] JP 2020-86778 A Summary of the Invention
[0004] The inventors came up with the idea of automating the updating of the learning model. In this case, it is conceivable that the learning model automatically acquires data for re-learning and re-learns at a specific timing such as periodically.
[0005] However, when data for re-learning is acquired automatically and re-learning is performed in this way, it is unclear whether the re-learning is appropriate. This is because the suitability of the data for re-learning is not verified, or even if it is verified, the verification is not necessarily valid. As a result, there is a possibility that the learning model will produce an inappropriate output for the input value. Even in such a case, it is desirable to be able to verify it after the fact.
[0006] A re-learning data management system according to one aspect of the present disclosure is a re-learning data management system configured to perform management processing for a first AI (Artificial Intelligence) system that executes inference processing using a machine learning model, wherein the first AI system is configured to obtain re-learning data for the machine learning model and re-learn the machine learning model, and the management processing comprises obtaining the re-learning data and storing the re-learning data as re-learning data in the first AI system in a database accessible from a verifier computer provided on a computer different from the first AI system.
[0007] A re-learning data management method according to one aspect of the present disclosure is a re-learning data management method for a first AI (Artificial Intelligence) system that performs inference processing using a machine learning model, wherein the first AI system is configured to obtain re-learning data for the machine learning model and re-learn the machine learning model, and the re-learning data management method comprises obtaining the re-learning data and storing the re-learning data as re-learning data in the first AI system in a database accessible from a verifier computer provided on a computer different from the first AI system.
[0008] Further details will be described in the following embodiments. [Brief description of the drawings]
[0009] [Figure 1] FIG. 1 is a configuration diagram of a data management system according to an embodiment. [Diagram 2] FIG. 2 is a configuration diagram of a computer network system included in the data management system. [Diagram 3] Figure 3 is a configuration diagram of an AI (Artificial Intelligence) system. [Figure 4] FIG. 4 is a diagram showing a flow of a data management method according to an embodiment. [Diagram 5]FIG. 5 is a flowchart showing the flow of processing in the AI system in the first embodiment. [Figure 6] FIG. 6 is a flowchart showing the processing flow of a smart contract installed in a computer network system. [Figure 7] FIG. 7 is a flowchart showing the processing flow of a smart contract installed in a computer network system. [Figure 8] FIG. 8 is a flowchart showing the processing flow of a smart contract installed in a computer network system. [Figure 9] FIG. 9 is a diagram for explaining a data management method according to the embodiment. [Figure 10] FIG. 10 is a flowchart showing the flow of processing in the AI system in the second embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0010] <1. Overview of the re-learning data management system and the re-learning data management method>
[0011] (1) A re-learning data management system according to an embodiment is a re-learning data management system configured to perform management processing for a first AI (Artificial Intelligence) system that executes inference processing using a machine learning model, in which the first AI system is configured to obtain re-learning data for the machine learning model and re-learn the machine learning model, and the management processing includes obtaining the re-learning data and storing the re-learning data as data for re-learning in the first AI system in a database accessible from a verifier computer provided on a computer different from the first AI system.
[0012] A machine learning model is trained using training data. Training data is data used for machine learning. Re-training refers to further training, using training data, on a trained machine learning model used in an AI system. Re-training data refers to training data used for re-training. Re-training updates a trained machine learning model used in an AI system.
[0013] The re-learning data may be selected from data used in an inference process using a machine learning model in the first AI system. As another example, the re-learning data may be provided from a device other than the first AI system.
[0014] The verifier computer is a device other than the first AI system, and refers to a computer used to verify a machine learning model used in the inference process of the first AI system. The verifier computer may be an AI system other than the first AI system (e.g., a second AI system).
[0015] Being accessible from the verifier computer means being able to obtain data from the database. As an example, the data may be obtained by the verifier computer requesting data from the database or a device that manages the database. As another example, the data may be obtained by the verifier computer directly communicating with a computer that includes the database.
[0016] This allows the re-learning data in the first AI system to be stored in a database accessible from the verifier computer, which is provided in a computer different from the first AI system. This allows the verifier computer to obtain the re-learning data. As a result, processing using the re-learning data, such as verification of the first AI system, becomes possible in other devices such as the verifier computer. By allowing the verifier computer to verify the first AI system, it becomes possible to verify after the fact whether the machine learning model is appropriate.
[0017] (2) Preferably, the database is configured as a distributed ledger using a distributed computer network system, which can increase the fairness of the re-learning data stored in the database.
[0018] (3) Preferably, the management process is a process realized by a computer constituting the distributed computer network system executing a smart contract implemented in the distributed computer network system.
[0019] A smart contract is software (computer program) executed by a computer that constitutes a blockchain. For example, a system administrator implements a smart contract in a blockchain so that the smart contract is executed in the blockchain. The use of a smart contract facilitates the process of storing re-learning data in a database.
[0020] In addition, since smart contracts may have restrictions on the amount of program code that can be written (the number of steps that can be written), when multiple processes are executed by a smart contract, it is easy to avoid an increase in the number of steps in a single process by separating each process.
[0021] (4) Preferably, the management process realized by the smart contract further comprises providing re-training data to the verifier computer in response to a request from the verifier computer, whereby the verifier computer can perform processes such as verification of the first AI system using the re-training data.
[0022] (5) Preferably, the management process realized by the smart contract further comprises storing attributes of the re-learning data, the attributes including at least one of identification information of the first AI system, identification information of a device that generated the re-learning data, identification information of a location where the re-learning data was generated, and identification information at the time of generation of the re-learning data. By storing such attributes, the re-learning data stored in the database can be managed using the attributes.
[0023] (6) Preferably, providing the verifier computer with the re-training data includes providing the verifier computer with the re-training data stored in the database that is identified by an attribute included in the request from the verifier computer. This makes it possible to provide the verifier computer with the re-training data requested by the verifier computer and also makes it easier to make requests from the verifier computer.
[0024] (7) Preferably, the management process realized by the smart contract further includes assigning a code to the re-learning data obtained from the first AI system and transmitting the code to the first AI system. The code may be stored in the first AI system. This makes it possible for the first AI system to prove that it has stored the re-learning data in a database so that it can be provided to the verifier computer.
[0025] (8) Preferably, the code is a code that permits the first AI system to re-learn using the re-learning data. In this way, in order for the first AI system to re-learn using the re-learning data, it is necessary to store the training data in a database so that it can be provided to the verifier computer. As a result, the re-learning data is stored in the database so that it can be provided to the verifier computer.
[0026] (9) Preferably, the management process realized by the smart contract further comprises notifying a second AI system, different from the first AI system, of the anomaly of the first AI system in response to obtaining a verification result of the anomaly of the first AI system from the verifier computer, whereby the second AI system can know the anomaly of the first AI system and can take measures.
[0027] (10) Preferably, notifying the anomaly includes transmitting information related to the re-learning data obtained from the first AI system to the second AI system, thereby enabling the second AI system to deal with the anomaly of the first AI system by using the information related to the re-learning data obtained from the first AI system.
[0028] (11) A method for managing re-learning data in accordance with an embodiment is a method for managing re-learning data for a first AI system that executes inference processing using a machine learning model, wherein the first AI system is configured to obtain re-learning data for the machine learning model and re-learn the machine learning model, and the method for managing re-learning data comprises obtaining the re-learning data and storing the re-learning data as data for re-learning in the first AI system in a database accessible from a verifier computer provided on a computer different from the first AI system.
[0029] This allows the re-learning data in the first AI system to be stored in a database accessible from the verifier computer, which is provided in a computer different from the first AI system. This allows the verifier computer to obtain the re-learning data. As a result, processing using the re-learning data, such as verification of the first AI system, becomes possible in other devices such as the verifier computer. By allowing the verifier computer to verify the first AI system, it becomes possible to verify after the fact whether the machine learning model is appropriate.
[0030] <2. Examples of re-learning data management systems and re-learning data management methods>
[0031] [First embodiment]
[0032] 1, a data management system 1 according to an embodiment includes a smart contract 121 implemented in a computer network system 100 such as a blockchain. As shown in FIG. 1, the blockchain is configured by a P2P computer network in which a plurality of computers are interconnected.
[0033] The smart contract 121 is software (computer program) implemented in the computer network system 100 such as a blockchain, and is configured to execute the data management process according to the present embodiment. The smart contract 121 can be executed by any of the multiple computers constituting the computer network system 100.
[0034] In the following description, the operation of any of the computers constituting the computer network system 100 that executes the smart contract 121 will be described as the operation of the smart contract 121. In the following description, one computer that executes the smart contract 121 among the multiple computers constituting the computer network system 100 will also be referred to as the computer network system 100.
[0035] The computer network system 100 is connected to an AI (Artificial Intelligence) system (first AI system) 300 via a network such as the Internet. The AI system 300 is a computer having a machine learning model 312 (FIG. 3).
[0036] The computer network system 100 may be connected to a first AI system 300, a second AI system 300B, ... an n-th AI system 300N, as shown in Fig. 1. In this case, in the following description, the AI system 300 refers to one AI system among the multiple AI systems 300, 300B, ... 300N. The AI system 300 and the smart contract 121 can communicate with each other.
[0037] The AI system 300 is equipped with a trained machine learning model 312. The AI system 300 performs inference processing 311 ( FIG. 3 ) using the machine learning model 312. As an example, the AI system 300 is a system that performs driving control of a vehicle 700 that performs autonomous driving. In this case, the inference processing 311 is processing for inferring driving control from information obtained while the vehicle 700 is traveling, and specifically includes processing for inferring the surrounding environment from images captured by the vehicle 700.
[0038] 1, the AI system 300 is a system that performs driving control of the vehicle 700, and the first AI system 300 performs driving control of the vehicles 700 and 700A. Similarly, the second AI system 300B performs driving control of the vehicle 700B.
[0039] The machine learning model 312 is trained using training data. Training data is data used for machine learning. Training refers to training the machine learning model so that it produces appropriate outputs. Machine learning adjusts the parameters of the model.
[0040] When the learning is supervised learning, the learning data is a pair of data consisting of an input and a corresponding output. In this example, the learning data may be, for example, a pair of data in which a captured image including a road sign is used as input data and the content indicated by the road sign is used as output data. Other types of learning include unsupervised learning and reinforcement learning, and the learning data may be data corresponding to each type of learning. The machine learning model 312 has parameters appropriately updated by repeating learning, thereby improving inference accuracy.
[0041] The AI system 300 obtains re-learning data and re-learns the trained machine learning model 312. Re-learning refers to further learning using the training data that is performed on the trained machine learning model 312 in the AI system 300. The training data refers to the training data used for re-learning. Through re-learning, the machine learning model 312 is updated and evolves.
[0042] The smart contract 121 acquires re-learning data D1, D2, ..., Dn used for re-learning of each of the first AI system 300, the second AI system 300B, ..., the n-th AI system 300N. One of the re-learning data D1, D2, ..., Dn is also referred to as re-learning data D.
[0043] The data management system 1 includes a relearning data ledger 122 implemented in the computer network system 100. The relearning data ledger 122 is a database (storage area) provided in a computer different from the AI system 300. In detail, the relearning data ledger 122 is a database possessed by each of the multiple computers constituting the computer network system 100. The databases possessed by each of the multiple computers are synchronized using distributed ledger technology.
[0044] The smart contract 121 stores the re-learning data D acquired from the AI system 300 in the re-learning data ledger 122. By storing the re-learning data D in the re-learning data ledger 122 held in the computer network system 100 such as a blockchain, the fairness of the re-learning data D stored in the re-learning data ledger 122 is improved.
[0045] The computer network system 100 is connected to a verifier computer 500 via a network such as the Internet. The verifier computer 500 is a computer different from the AI system 300 to be verified (e.g., the first AI system 300), and is a computer used by a verifier who performs verification (described below) using the re-learning data of the first AI system 300 stored in the re-learning data ledger 122. The verifier computer 500 may include an AI system 300 (e.g., the second AI system 300B) other than the first AI system 300 to be verified.
[0046] The relearning data ledger 122 is a database accessible from the verifier computer 500. Being accessible means that data can be obtained from the relearning data ledger 122. As an example, data may be obtained by the verifier computer 500 requesting data from the smart contract 121 and obtaining the data from the smart contract 121, as shown in FIG. 1. As another example, data may be obtained by the verifier computer 500 directly communicating with the relearning data ledger 122 to obtain the data.
[0047] FIG. 2 is a block diagram showing the schematic configuration of each of the multiple computers that make up the computer network system 100, and shows the schematic configuration of the computer network system 100 as one of the computers.
[0048] 2, the computer network system 100 includes a processor 11 and a memory 12. The memory 12 is connected to the processor 11. The memory 12 includes, for example, a primary storage device and a secondary storage device. The primary storage device is, for example, a RAM. The secondary storage device is, for example, a hard disk drive (HDD) or a solid state drive (SSD).
[0049] The memory 12 includes a smart contract 121 as a computer program executed by the processor 11. The processor 11 reads and executes the smart contract 121 stored in the memory 12. The smart contract 121 has program code for processing executed by the computer network system 100, and may include one or more program modules each of which causes the processor 11 to execute processing.
[0050] The memory 12 includes a relearning data ledger 22 as a storage area for storing relearning data by the smart contract 121. The memory 12 also includes an attribute storage unit 123 as a storage area for storing attributes of the relearning data, which will be described later. The attribute storage unit 123 stores attribute data 124 indicating attributes of each piece of relearning data. By storing the attributes, the relearning data stored in the relearning data ledger 22 can be managed using the attributes.
[0051] By reading and executing the smart contract 121, the processor 11 executes the storage process 111. That is, the smart contract 121 executes the storage process 111. The storage process 111 includes acquiring re-learning data D1, D2, ... Dn in each of the AI systems 300 and storing the data in the re-learning data ledger 122. As an example, it is assumed that the re-learning data D1, D2, ... Dn are received from each of the AI systems 300.
[0052] The storage process 111 includes an extraction process 112. The extraction process 112 includes extracting attributes from the re-learning data D received from the AI system 300. The extracted attributes include at least one of identification information (e.g., ID (identification)) of the AI system 300, identification information of the device that generated the re-learning data D, identification information of the location where the re-learning data D was generated, and identification information at the time of generation of the re-learning data D.
[0053] The identification information of the AI system 300 can be obtained by analyzing the source of the re-learning data D. When the AI system 300 uses, as input data, captured images obtained from the vehicle 700 as the re-learning data D, the identification information of the device that generated the re-learning data D becomes the identification information (vehicle ID) of the vehicle 700, and can be obtained by analyzing the source of the captured images used as the re-learning data D in the AI system 300.
[0054] In the above example, the identification information at the time of generating the re-learning data D is identification information at the time when the vehicle 700 captured an image, for example, the date and time of the image capture. In the above example, the identification information of the generation location of the re-learning data D is identification information of the location where the vehicle 700 captured an image, for example, the location information of the image capture location. These can be obtained by analyzing the captured image included in the re-learning data D.
[0055] In detail, the storage process 111 includes a code assignment process 113. The code assignment process 113 includes assigning a code to the re-learning data D obtained from the AI system 300. The code refers to identification information unique to the re-learning data.
[0056] The code assignment process 113 includes transmitting the code assigned to the re-learning data D to the AI system 300, which is the sender of the re-learning data. By obtaining this code, the AI system 300 can use it as proof that the re-learning data D has been registered in the computer network system 100.
[0057] In some aspects, registering the re-learning data D in the computer network system 100 serves as proof that the machine learning model 312 has been re-learned. Therefore, the acquisition of the code can serve as proof that the machine learning model 312 has been updated in the AI system 300. In addition, as described below, the re-learning data D can also be read by the verifier computer 500 and used to verify the machine learning model 312 in the AI system 300. Therefore, the acquisition of the code can serve as proof that the AI system 300 is an AI system capable of verifying the machine learning model 312.
[0058] Preferably, the storage process 111 associates a code with the re-learning data D from the AI system 300 and stores the code in the re-learning data ledger 122. This makes it possible to easily identify the re-learning data D stored in the re-learning data ledger 122 using the code.
[0059] The storage process 111 includes associating a code with the attribute extracted from the re-learning data D and storing the code as attribute data 124 in the attribute storage unit 123. This allows the code to be identified from the attribute, and the re-learning data D to be easily identified from the code.
[0060] The smart contract 121 executes the verification process 114. The verification process 114 includes providing the re-learning data D to the verifier computer 500 in response to a request from the verifier computer 500.
[0061] Providing the re-learning data D to the verifier computer 500 may, for example, involve reading the re-learning data D corresponding to the request from the re-learning data ledger 122 and transmitting it to the verifier computer 500. In this way, the verifier computer 500 can obtain the requested re-learning data D.
[0062] At this time, the smart contract 121 identifies the re-learning data D based on the request received from the verifier computer 500. Therefore, preferably, the verification process 114 includes a read process 115. The read process 115 includes reading attributes of the re-learning data D based on the request received from the verifier computer 500.
[0063] One example of the attribute read from the request is the identification information of the AI system 300. This is assumed to be the case where the verifier computer 500 specifies the identification information of the AI system 300 to be verified and requests the smart contract 121 for re-learning data D.
[0064] Another example of the attribute read from the request is the identification information of the vehicle 700 that generated the re-learning data D. This is assumed to be the case where the verifier computer 500 requests the re-learning data D from the smart contract 121 by specifying the identification information of the vehicle 700 that is the generator of the re-learning data D to be verified.
[0065] Other examples of the attributes read from the request include identification information of the location where the re-learning data D was generated and identification information at the time of generation of the re-learning data D. This is assumed to be the case where the verifier computer 500 specifies the photographing location and photographing time of the re-learning data D to be verified and requests the re-learning data D from the smart contract 121.
[0066] Identifying the re-learning data D includes, for example, identifying a code corresponding to the attribute read from the request by referring to the attribute data 124 in the attribute storage unit 123, and reading the re-learning data D corresponding to the code from the re-learning data ledger 122. Identifying the re-learning data D may, for example, be searching the re-learning data ledger 122 using the attribute read from the request, and reading the re-learning data D corresponding to the attribute read from the request from the re-learning data ledger 122. This allows the smart contract 121 to transmit the re-learning data D requested by the verifier computer 500 to the verifier computer 500. Furthermore, the verifier computer 500 can request the re-learning data D with the attribute according to the verification from the smart contract 121.
[0067] The smart contract 121 executes a notification process 116. The notification process 116 includes notifying a device other than the first AI system 300 of the verification result of the first AI system 300 obtained from the verifier computer 500. The verification result of the first AI system 300 may be normal, abnormal, or both. It may also only be that verification has been performed. This allows a device other than the first AI system 300 to know the verification result of the first AI system 300 and that verification has been performed.
[0068] The device different from the first AI system 300 may be, for example, the other AI systems 300B, ..., 300N. This allows the other AI systems 300B, ..., 300N to understand the verification results of the first AI system 300 and the fact that verification has been performed.
[0069] As one example, the notification process 116 is performed in response to obtaining a verification result of the first AI system 300 from the verifier computer 500. The notification process 116 may be performed, for example, when a specific verification result such as abnormal or normal is obtained. As another example, the notification process 116 may be performed a predetermined number of times when a verification result is received from the verifier computer 500. This allows the verification result of the first AI system 300 to be known at an appropriate time.
[0070] The notifying includes transmitting to another device information related to the re-learning data D obtained from the first AI system 300 and related to the verification. The information related to the re-learning data D may be a code of the re-learning data D, an attribute of the re-learning data D, the re-learning data D itself, or a combination thereof.
[0071] Furthermore, the re-learning data D related to the verification is, for example, re-learning data D related to the timing notified from the verifier computer 500 together with the verification result, and examples include re-learning data D including images captured by the vehicle 700 at the notified timing, re-learning data D transmitted from the first AI system 300 at the notified timing, etc. As other examples, the re-learning data D may be re-learning data D including images captured by the vehicle 700 during a predetermined period prior to the notified timing, re-learning data D transmitted from the first AI system 300, etc.
[0072] For example, when an abnormal verification result is notified from the verifier computer 500, the re-learning data D that may have caused the abnormality is notified to the other devices. When the other devices are the other AI systems 300B, ..., 300N, it is possible to prevent re-learning using the notified re-learning data D or re-learning data related to the re-learning data D in those AI systems 300. As a result, it is possible to prevent the occurrence of an abnormality in the other AI systems 300B, ..., 300N. In addition, the occurrence of an abnormality can be predicted by checking whether or not re-learning has been performed using the notified re-learning data D or re-learning data related to the re-learning data D in the other AI systems 300B, ..., 300N.
[0073] 3, the AI system 300 is assumed to be a general computer including a processor 31 and a memory 32. The memory 32 is connected to the processor 31. The memory 32 includes, for example, a primary storage device and a secondary storage device.
[0074] The memory 32 includes an inference program 321 as a computer program executed by the processor 31. The processor 31 executes the inference process 311 by reading the inference program 321 from the memory 32 and executing it.
[0075] The inference process 311 includes inputting data to the machine learning model 312 and acquiring data output from the machine learning model 312. As an example, the AI system 300 infers road signs installed in the traveling direction of the vehicle 700 from an image captured by the vehicle 700. In this case, in the inference process 311, the processor 31 uses the image captured by the vehicle 700 as input data for the machine learning model 312 and acquires labels indicating road signs output from the machine learning model 312.
[0076] The inference process 311 includes a first transmission process 313 as an example of a process using the inference result. The first transmission process 313 includes transmitting the inference result to a predetermined output destination. In this example, the first transmission process 313 includes transmitting a control signal generated based on the label output from the machine learning model 312 to the vehicle 700. In this way, the driving control of the vehicle 700 is realized using the machine learning model 312.
[0077] The inference process 311 preferably includes a storage process 314. The storage process 314 includes storing the inference result in the inference process 311 as inference result data 325 in an inference result storage unit 324, which is a storage area of the memory 32. The inference result data 325 includes a combination of an image captured by the vehicle 700, which is input data to the machine learning model 312, and a label indicating a road sign output from the machine learning model 312. By storing the inference result data 325 in the memory 32, the inference result data 325 can be used as re-learning data in the re-learning process 315 described later.
[0078] The memory 32 includes a re-learning program 322 as a computer program executed by the processor 31. The processor 31 executes the re-learning process 315 by reading the re-learning program 322 from the memory 32 and executing it.
[0079] The re-learning process 315 includes re-learning the machine learning model 312. In the re-learning process 315, re-learning data is used. As an example of the re-learning data, a combination of a captured image and a label included in the inference result data 325 stored in the inference result storage unit 324 is used. This makes it possible to easily obtain the re-learning data. Preferably, a combination of a captured image and a label determined to be suitable for re-learning, among the inference result data 325 stored in the inference result storage unit 324, is used. This makes it possible to easily obtain re-learning data suitable for re-learning. As an example, the determination of whether or not the inference result is suitable for re-learning may be input from an external device.
[0080] In the above example, the re-learning process 315 includes a selection process 316. The selection process 316 includes selecting a combination of a captured image and a label to be used as re-learning data from the inference result data 325 stored in the inference result storage unit 324 in accordance with a predetermined condition. In the above example, the predetermined condition is, for example, in accordance with the contents of an input from an external device or the like. In another example, the predetermined condition may be a combination of a captured image and a label used in the inference process 311 at a predetermined timing. In this way, re-learning data is obtained.
[0081] It is to be noted that the second code process 319B included in the re-learning process 315 shown in FIG. 3 is used in the second embodiment.
[0082] The memory 32 includes a recording program 323 as a computer program executed by the processor 31. The processor 31 executes the recording process 317 by reading the recording program 323 from the memory 32 and executing it.
[0083] The recording process 317 includes a second transmission process 318, and includes passing the relearning data D used in the relearning process 315 to the communication device 33 and having it transmitted to the smart contract 121. As a result, the relearning data D is transmitted to the smart contract 121 and recorded in the relearning data ledger 22.
[0084] Preferably, the recording process 317 includes a first code process 319A. The first code process 319A includes receiving a code assigned to the transmitted re-learning data D from the smart contract 121, and adding the code to the inference result data 325 in association with a combination of a captured image and a label corresponding to the re-learning data D of the inference result data 325.
[0085] 3, a code is assigned to the combination of the captured image and the label recorded in the relearning data ledger 22 as the relearning data D in the inference result data 325. The inclusion of a code in the inference result data 325 indicates that the relearning in the AI system 300 can be verified by the verifier computer 500.
[0086] A data management method according to the embodiment will be described with reference to Figures 4 and 9. The data management method according to the embodiment is based on the premise that in the first AI system 300, re-learning (step S30) of the machine learning model 312 is performed at an appropriate timing during the automatic driving control (step S10) of the vehicle 700.
[0087] In the automatic driving control of step S10, the vehicle 700 captures an image of the traveling direction, etc. (step S11). As shown in Fig. 9, in step S11, if a road sign 5 is included in the capture range, a captured image 60 including an image of the road sign 5 is obtained.
[0088] When the vehicle 700 transmits image data of the captured image 60 to the first AI system 300 (step S13), the first AI system 300 infers a sign from the captured image (step S15).
[0089] The first AI system 300 transmits a control signal SG based on the inference result to the vehicle 700 (step S17). In the vehicle 700, driving control is realized based on the control signal (step S19). In this way, the automatic driving of the vehicle 700 is realized.
[0090] The data management method according to the embodiment includes registering re-learning data D (step S50), which is performed at an appropriate timing between the automatic driving in step S10 and the re-learning in step S20. In step S50, the first AI system 300 that performed the re-learning in step S20 registers the re-learning data D in the memory 32 (step S51) and transmits the re-learning data D to the smart contract 121 (step S53). The smart contract 121 registers the re-learning data D in the re-learning data ledger 122 (step S55).
[0091] The data management method according to the embodiment includes providing the verifier computer 500 with the relearning data D registered in the relearning data ledger 122 (step S70). The processing of step S70 is performed at an appropriate timing between the automatic driving of step S10 and the relearning of step S20. As an example, step S70 is started when the verifier computer 500 transmits a request for the relearning data D to the smart contract 121 (step S71).
[0092] In the processing of step S70, the smart contract 121 reads the re-learning data D in response to the request of step S71 from the re-learning data ledger 122 (step S73) and transmits it to the verifier computer 500 (step S75). This makes it possible for the verifier computer 500 or the like to verify whether the machine learning of the machine learning model 312 is normal or abnormal.
[0093] The data management method according to the embodiment includes a process (step S90) of sending a notification in response to a verification result from the verifier computer 500. The process of step S90 is performed at an appropriate timing between the automatic driving of step S10 and the relearning of step S20. As an example, the process of step S90 is started when the verification result is transmitted from the verifier computer 500 to the smart contract 121 (step S91).
[0094] In step S91, it is assumed that the verification result of the first AI system 300 is transmitted from the verifier computer 500. In this case, the smart contract 121 reads out the re-learning data D related to the verification result of the first AI system 300 from the re-learning data ledger 122 (step S93). Then, the smart contract 121 notifies an AI system other than the first AI system 300, for example, the second AI system 300B, based on the verification result (step S95).
[0095] In the AI system 300, as an example, processing is performed according to the flow shown in Fig. 5. That is, referring to Fig. 5, when image data is received from the vehicle 700 (YES in step S101), the processor 31 of the AI system 300 inputs the received image data to the machine learning model 312 (step S103) and receives a label indicating a road sign from the machine learning model 312 (step S105). The processor 31 stores the image data input to the machine learning model 312 and the label output from the machine learning model 312 as an inference result in the memory 32 (step S107).
[0096] The processor 31 obtains a control signal by generating a control signal for driving control of the vehicle 700 from the obtained inference result (step S109), and transmits the control signal to the vehicle 700 (step S111). As a result, the driving of the vehicle 700 is controlled.
[0097] When a predetermined timing is reached (YES in step S113), the processor 31 selects from the memory 32 the inference result to be used as re-learning data D for the machine learning model 312 (step S115), and re-learns the machine learning model 312 (step S117). The processor 31 transmits the re-learning data D used for the re-learning to the smart contract 121 (step S119). As a result, the re-learning data D is registered in the re-learning data ledger 122.
[0098] After transmitting the re-learning data D in step S119, the processor 31 receives the code assigned to the re-learning data D transmitted in step S119 from the smart contract 121 (step S121). The processor 31 stores the code obtained from the smart contract 121 in step S121 in the memory 32 in association with the inference result used as the re-learning data D (step S123).
[0099] By repeatedly executing the above processes, the processor 31 re-learns the machine learning model 312 while performing processing using the machine learning model 312, and registers the re-learning data D in the re-learning data ledger 122. The re-learning data D stored in the re-learning data ledger 122 can be provided to the verifier computer 500 by the smart contract 121, that is, it is stored in the re-learning data ledger 122 that can be accessed from devices other than the AI system 300. This makes it possible to perform verification using the re-learning data D by devices other than the AI system 300.
[0100] As an example, the smart contract 121 performs a process of registering the re-learning data D used for re-learning the machine learning model 312 in the AI system 300, in the flow shown in Fig. 6. That is, referring to Fig. 6, when the smart contract 121 receives the re-learning data D from the AI system 300 (YES in step S201), it extracts attributes (step S203). In addition, the smart contract 121 assigns a code (step S205).
[0101] The smart contract 121 registers the re-learning data D together with the code in the re-learning data ledger 122 (step S207). The smart contract 121 also stores the attribute extracted in step S203 together with the code in the memory 12 (step S209).
[0102] By repeatedly executing the above processes by the processor 11, the re-learning data D used for re-learning the machine learning model 312 in the AI system 300 is registered in the re-learning data ledger 122. In other words, the re-learning data D used for re-learning the machine learning model 312 in the AI system 300 can be made in a state where verification, etc. can be performed using the re-learning data D by a device other than the AI system 300.
[0103] As an example, the smart contract 121 performs a process of providing the re-learning data D used in the AI system 300 to re-learn the machine learning model 312 to another device, following the flow shown in Fig. 7. That is, referring to Fig. 7, when the smart contract 121 receives a request for the re-learning data D from the verifier computer 500 (YES in step S301), it extracts attributes of the re-learning data D to be sent from the request (step S303).
[0104] The smart contract 121 identifies the code corresponding to the attribute extracted in step S303, and reads the re-learning data D to which the code is assigned from the re-learning data ledger 122 (step S305). The smart contract 121 transmits the re-learning data D read from the re-learning data ledger 122 to the verifier computer 500 (step S307).
[0105] By repeatedly executing the above processes by the processor 11, the re-learning data D used in the AI system 300 to re-learn the machine learning model 312 is provided to a device other than the AI system 300, such as the verifier computer 500. In other words, the re-learning data D used in the AI system 300 to re-learn the machine learning model 312 can be provided so that a device other than the AI system 300 can perform verification using the re-learning data D.
[0106] As an example, the smart contract 121 performs a process of notifying other devices based on the verification result of the AI system 300 by the verifier computer 500 or the like, in the flow shown in Figure 8. That is, referring to Figure 8, when the smart contract 121 receives the verification result of the AI system 300 from the verifier computer 500 (YES in step S401), it extracts attributes of the re-learning data D to be used for notification from the verification result (step S403).
[0107] The smart contract 121 identifies the code corresponding to the attribute extracted in step S403, and reads the re-learning data D to which the code is assigned from the re-learning data ledger 122 (step S405). The smart contract 121 transmits the re-learning data D read from the re-learning data ledger 122 to another device, such as the second AI system 300B, different from the AI system 300 (step S407).
[0108] When the processor 11 repeatedly executes the above process and obtains a verification result indicating that an abnormality has occurred in the machine learning model 312, the re-learning data D of the AI system 300 related to the verification is transmitted to another device such as the second AI system 300B. This notifies the other device of the verification result. In other words, the verification result, such as abnormality or normality of the machine learning model 312 of the AI system 300, can be utilized in other AI systems, etc.
[0109] For example, when a captured image 60 including an image of the road sign 5 shown in FIG. 9 is transmitted from the vehicle 700 to the AI system 300, the AI system 300 performs inference on the road sign 5 using the machine learning model 312.
[0110] The road sign 5 may differ from the original display, for example, by having dirt 51 attached thereto, as shown in Fig. 9. The difference from the original display may be accidental, as in the example of Fig. 9, artificial, or a combination thereof.
[0111] In the example of FIG. 9, road sign 5 is a sign indicating a speed limit, and indicates that the speed limit is 60 kilometers per hour. However, dirt 51 is attached to road sign 5, which indicates that 6 is displayed closer to 8. In this case, it is assumed that the label output from machine learning model 312 will be label 61, which is "drive at 60 kilometers per hour or less," or label 62, which is "drive at 80 kilometers per hour or less."
[0112] In the former case, the machine learning model 312 is normal, and a control signal SG based on the label 61 is transmitted to the vehicle 700. This causes the vehicle 700 to travel in accordance with the road signs 5.
[0113] On the other hand, in the latter case, the machine learning model 312 is abnormal, and a control signal SG based on the label 62 is sent to the vehicle 700. This may cause the vehicle 700 to travel at a speed greater than the speed limit of 60 kilometers per hour, even though the road sign 5 indicates that the speed limit is 60 kilometers per hour.
[0114] When the machine learning model 312 is re-trained using the captured image 60 and the label 61, re-learning data D including the captured image 60 and the label 61 is transmitted to the smart contract 121 and registered in the re-learning data ledger 122. In this case, the machine learning model 312 is re-trained appropriately.
[0115] When the machine learning model 312 is re-trained using the captured image 60 and the label 62, re-learning data D including the captured image 60 and the label 62 is transmitted to the smart contract 121 and registered in the re-learning data ledger 122. When the machine learning model 312 is re-trained using the re-learning data D in this case, the machine learning model 312 is updated in a way that degrades the inference accuracy of the machine learning model 312.
[0116] By using the re-learning data D registered in the re-learning data ledger 122, an anomaly in the machine learning model 312 that has occurred in this manner can be verified in another device such as the verifier computer 500. In this case, as an example, the re-learning data D including the captured image 60 and the label 62 is transmitted to another device such as the second AI system 300B. This can be used to check whether re-learning has been performed using similar re-learning data, to prevent re-learning from being performed using similar re-learning data in the future, and to detect or prevent anomalies in other systems 300.
[0117] [Second embodiment]
[0118] The code provided by the smart contract 121 may be used as an authorization code for re-learning. In the second embodiment, when the AI system 300 performs the selection process 316 to select the re-learning data D for re-learning the machine learning model 312, the AI system 300 performs the recording process 317 to transmit the re-learning data D to the smart contract 121.
[0119] In the second embodiment, the re-learning process 315 of the AI system 300 includes a second code process 319B shown in Fig. 3. The second code process 319B includes using a code provided by the smart contract 121 as an authorization code, and authorizing the re-learning process 315 for re-learning the machine learning model 312 when the authorization code is provided to the re-learning data D.
[0120] In the second embodiment, the AI system 300 executes, as an example, the process flow shown in Fig. 10. In Fig. 10, the processes with the same step numbers as those in the flowchart of Fig. 5 are the same as the processes in Fig. 5.
[0121] 10, when a predetermined timing is reached (YES in step S113), the processor 31 of the AI system 300 selects an inference result to be used as re-learning data D of the machine learning model 312 from the memory 32 (step S115). At this time, in the second embodiment, the processor 31 transmits the selected re-learning data D to the smart contract 121 (step S513). As a result, the re-learning data D is registered in the re-learning data ledger 122.
[0122] When the processor 31 re-learns the machine learning model 312 using the re-learning data D, the processor 31 checks whether or not a code has been assigned to the re-learning data D. If a code has been assigned, that is, if the code assigned to the re-learning data D transmitted in step S513 has been received from the smart contract 121 (YES in step S515), the processor 31 re-learns the machine learning model 312 using the re-learning data D (step S517).
[0123] On the other hand, if a code has not been assigned to the re-learning data D, that is, if a code assigned to the re-learning data D has not been received from the smart contract 121 (NO in step S515), the processor 31 skips step S517.
[0124] By repeatedly executing the above processes by the processor 31, the re-learning of the machine learning model 312 is performed using only the re-learning data D transmitted to the smart contract 121, and re-learning is not performed using the re-learning data D not transmitted to the smart contract 121. Therefore, in the AI system 300, in order to re-learn the machine learning model 312, it is necessary to transmit the re-learning data D to be used to the AI system 300 before re-learning. As a result, all of the re-learning data D used for re-learning in the AI system 300 is registered in the re-learning data ledger 122. By utilizing the code in this way, it becomes easier to verify the machine learning model 312 of the AI system 300.
[0125] <3. Notes> The present invention is not limited to the above-described embodiment, and various modifications are possible. [Explanation of symbols]
[0126] 1: Data management system 5: Road sign 11: Processor 12: Memory 22: Re-learning data ledger 31: Processor 32: Memory 33: Communication equipment 60: Photographed image 61: Label 62: Label 100: Computer network systems 111: Storage process 112: Extraction process 113: Code assignment process 114: Verification process 115: Read process 116: Notification processing 121: Smart Contracts 122: Re-learning data ledger 123: Attribute storage section 124: Attribute data 300: AI system 300B: Second AI system 300N: nth AI system 311: Inference processing 312: Machine learning models 313: First transmission process 314: Storage process 315: Re-learning process 316: Selection process 317: Recording Processing 318: Second transmission process 319A: First code processing 319B: Second code processing 321: Inference program 322: Relearning program 323: Recording program 324: Inference result storage unit 325: Inference result data 500: Verifier computer 700: Vehicle 700A: Vehicle 700B: Vehicle D: Re-learning data D1: Re-learning data SG: Control signal
Claims
1. A re-learning data management system configured to perform a management process for a first AI (Artificial Intelligence) system that performs an inference process using a machine learning model, The first AI system is configured to obtain data for retraining the machine learning model and retrain the machine learning model; The management process includes: Acquire the re-learning data; The re-learning data is stored as re-learning data in the first AI system in a database accessible from a verifier computer provided in a computer different from the first AI system; In response to obtaining a verification result of the abnormality of the first AI system from the verifier computer, notifying a second AI system different from the first AI system of the abnormality. Data management system for re-learning.
2. The database is configured as a distributed ledger using a distributed computer network system. The re-learning data management system according to claim 1.
3. The management process is realized by a computer constituting the distributed computer network system executing a smart contract implemented in the distributed computer network system. The re-learning data management system according to claim 2 .
4. The management process realized by the smart contract is providing the re-learning data to the verifier computer in response to a request from the verifier computer. The re-learning data management system according to claim 3.
5. The management process realized by the smart contract further comprises storing attributes of the re-learning data, The attribute includes at least one of identification information of the first AI system, identification information of a device that generated the re-learning data, identification information of a location where the re-learning data was generated, and identification information at the time of generation of the re-learning data. The re-learning data management system according to claim 4.
6. Providing the relearning data to the verifier computer includes providing the relearning data, among the relearning data stored in the database, to the verifier computer, the relearning data being identified by the attribute included in the request from the verifier computer. The re-learning data management system according to claim 5.
7. The management process realized by the smart contract is Assigning a code to the re-learning data obtained from the first AI system; transmitting the code to the first AI system. The re-learning data management system according to any one of claims 3 to 6.
8. The code is a code that permits the first AI system to re-learn using the re-learning data. The re-learning data management system according to claim 7.
9. A re-learning data management system configured to perform management processing for a first AI (Artificial Intelligence) system that performs inference processing using a machine learning model, comprising: The first AI system is configured to obtain data for retraining the machine learning model and retrain the machine learning model; The management process includes: Acquire the re-learning data; storing the re-learning data as re-learning data in the first AI system in a database provided in a computer different from the first AI system and accessible from a verifier computer of the first AI system; The database is configured as a distributed ledger using a distributed computer network system, The management process is realized by a computer constituting the distributed computer network system executing a smart contract implemented in the distributed computer network system, The management process realized by the smart contract is In response to obtaining a verification result of the abnormality of the first AI system from the verifier computer, notifying a second AI system different from the first AI system of the abnormality of the first AI system. Data management system for re-learning.
10. Notifying the anomaly includes transmitting information related to the re-learning data obtained from the first AI system to the second AI system. The re-learning data management system according to claim 9.
11. A re-learning data management method for a first AI system that executes inference processing using a machine learning model, comprising: The first AI system is configured to obtain data for retraining the machine learning model and retrain the machine learning model; The re-learning data management method includes: Acquire the re-learning data; The re-learning data is stored as re-learning data in the first AI system in a database accessible from a verifier computer provided in a computer different from the first AI system; In response to obtaining a verification result of the abnormality of the first AI system from the verifier computer, notifying a second AI system different from the first AI system of the abnormality. Data management method for retraining.
Citation Information
Patent Citations
Search device, search method, and plasma processing device
JP2019165123A
Machine learning model construction device and machine learning model construction method
JP2020086778A
Analysis support device and analysis support method
JP2020144504A
Information processing apparatus, information processing method and program
JP2020166418A
Security layer for configuring blockchain
WO2021013499A1
Cited By
Sensor module and information processing device
WO2026133549A1
