Model inference program, information processing device and model inference method

Abstract

To provide a model estimation program, an information processing apparatus, and a model estimation method that allow easy generation of a learning model.SOLUTION: A model estimation program accepts an estimation request of a model structure of a learning model required for generating the learning model and data features of learning data, refers to a storage unit that stores ontology including a hierarchy structure of each class, identifies an inference result including the model structure and the data features corresponding to the estimation request, and outputs the identified inference result.SELECTED DRAWING: Figure 6

Description

[Technical Field]

[0001] The present invention relates to a model inference program, an information processing device, and a model inference method. [Background technology]

[0002] In recent years, with the improvement in the performance of artificial intelligence (AI), the number of users (hereinafter simply referred to as "users") who utilize AI in their own businesses is increasing. Specifically, to utilize AI in their businesses, users generate and use learning models (e.g., DL: Deep Learning) that are tailored to their own businesses.

[0003] In this case, the user, for example, identifies a model structure that can realize his / her requirements and prepares the necessary training data.The user then performs training on the training data in the identified model structure to generate a training model with optimized parameters. [Prior art documents] [Non-patent literature]

[0004] [Non-Patent Document 1] Newton Howard, Naima Chouikhi, Ahsan Adeel, Katelyn Dial, Adam Howard and Amir Hussain: “BrainOS: A Novel Artificial Brain-Alike Automatic Machine Learning Framework”, https: / / doi.org / 10.3389 / fncom.2020.00016 [Non-patent document 2] Giuseppe Futia and Antonio Vetro: “On the Integration of Knowledge Graphs into Deep Learning Models for a More Comprehensible AI-Three Challenges for Future Research”, https: / / doi.org / 10.3390 / info11020122 Summary of the Invention [Problem to be solved by the invention]

[0005] However, such users often lack sufficient knowledge about AI and may be unable to generate a learning model by themselves. Specifically, users may be unable to, for example, prepare the learning data required to generate a learning model suited to their business or determine the structure of the learning model (hereinafter also referred to as model structure) suitable for generating the learning model by themselves. Therefore, when generating a learning model, the users may need to rely on AI experts, etc., and may not be able to easily generate the learning model.

[0006] Furthermore, much of the research on AI conducted in recent years has been aimed at optimizing the parameters of already-generated learning models, and there has not been sufficient research into methods for supporting users who do not have sufficient knowledge of AI in generating learning models.

[0007] SUMMARY OF THE INVENTION It is therefore an object of the present invention to provide a model inference program, an information processing device, and a model inference method that enable easy generation of a learning model. [Means for solving the problem]

[0008] To achieve the above-mentioned object, the model inference program of the present invention is characterized in that it causes a computer to execute a process of receiving an inference request for the model structure of a learning model and the data features of learning data required to generate a learning model, referring to a memory unit that stores an ontology having a hierarchical structure of each class, identifying an inference result including the model structure and the data features that correspond to the inference request, and outputting the identified inference result.

[0009] In addition, in one aspect, the model inference program of the present invention for achieving the above-mentioned object is characterized in that the ontology includes a first class corresponding to each of the items included in the inference request for each learning model, a second class corresponding to the task performed by each learning model, a third class corresponding to the model structure of each learning model, a fourth class corresponding to the data features of each learning model, and a fifth class corresponding to each of the items included in the inference result for each learning model.

[0010] In addition, in one aspect, the model inference program of the present invention for achieving the above-mentioned object is characterized in that the first class has a first subclass corresponding to each item regarding the function of each learning model, and a second subclass corresponding to each item other than the function of each learning model.

[0011] In addition, in one aspect, the model inference program of the present invention for achieving the above-mentioned object is characterized in that it generates a rule corresponding to the received inference request, references the memory unit that stores the ontology, and identifies the inference result corresponding to the generated rule.

[0012] In addition, in one aspect, a model inference program of the present invention for achieving the above-mentioned object is characterized in that it identifies a first instance corresponding to the received inference request from among the instances defined in the first subclass, identifies a second instance corresponding to the identified first instance from among the instances defined in the second class, identifies a third instance corresponding to the identified second instance from among the instances defined in the third class, and generates a first rule indicating that the model structure indicated by the identified third instance is to be identified as at least a part of the inference result.

[0013] In addition, in one aspect, the model inference program of the present invention for achieving the above-mentioned object is characterized in that it identifies a fourth instance corresponding to the received inference request from among the instances defined in the second subclass, and generates the first rule indicating that an instance corresponding to the identified second instance and fourth instance from among the instances defined in the third class is to be identified as the third instance.

[0014] In addition, in one aspect, the model inference program of the present invention for achieving the above-mentioned object is characterized in that it identifies a fifth instance corresponding to the identified third instance from among the instances defined in the fourth class, and generates the first rule indicating that the data feature indicated by the identified fifth instance is to be identified as at least part of the inference result.

[0015] In addition, in one aspect, the model inference program of the present invention for achieving the above-mentioned object is characterized in that it accepts multiple pieces of training data to be used in generating the learning model, and generates a second rule that indicates that the accepted multiple pieces of training data should be converted to correspond to the data features indicated by the identified fifth instance.

[0016] In addition, the information processing device of the present invention for achieving the above-mentioned object is characterized by having a request receiving unit that receives an inference request for the model structure of the learning model required for generating a learning model and the data features of the learning data, an inference execution unit that refers to a memory unit that stores an ontology having a hierarchical structure of each class and identifies an inference result including the model structure and the data features that correspond to the inference request, and a data output unit that outputs the identified inference result.

[0017] In addition, the model inference method of the present invention for achieving the above-mentioned object is characterized in that it has a computer execute a process to receive an inference request for the model structure of a learning model and the data features of learning data required to generate a learning model, refer to a memory unit that stores an ontology having a hierarchical structure for each class, identify an inference result including the model structure and the data features that correspond to the inference request, and output the identified inference result. [Effects of the Invention]

[0018] According to the model inference program, information processing device, and model inference method of the present invention, it becomes possible to easily generate a learning model. [Brief explanation of the drawings]

[0019] [Figure 1] FIG. 1 is a diagram illustrating the configuration of an information processing system 10. As shown in FIG. [Figure 2] FIG. 2 is a diagram illustrating the configuration of the information processing device 1. [Figure 3] FIG. 3 is a diagram illustrating the model inference process according to the first embodiment. [Figure 4] FIG. 4 is a diagram illustrating the model inference process according to the first embodiment. [Figure 5] FIG. 5 is a diagram illustrating details of the model inference process in the first embodiment. [Figure 6] FIG. 6 is a diagram illustrating details of the model inference process in the first embodiment. [Figure 7] FIG. 7 is a diagram illustrating details of the model inference process in the first embodiment. [Figure 8] FIG. 8 is a diagram illustrating details of the model inference process in the first embodiment. [Figure 9] FIG. 9 is a diagram illustrating details of the model inference process in the first embodiment. [Figure 10] FIG. 10 is a diagram illustrating details of the model inference process in the first embodiment. [Figure 11] FIG. 11 is a diagram illustrating details of the model inference process in the first embodiment. [Figure 12] FIG. 12 is a diagram illustrating details of the model inference process according to the first embodiment. [Figure 13] FIG. 13 is a diagram illustrating details of the model inference process according to the first embodiment. [Figure 14] FIG. 14 is a diagram illustrating details of the model inference process according to the first embodiment. [Figure 15] FIG. 15 is a diagram illustrating details of the model inference process in the first embodiment. [Figure 16] FIG. 16 is a diagram illustrating details of the model inference process according to the first embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0020] DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. However, the technical scope of the present invention is not limited to these preferred embodiments.

[0021] [Configuration of information processing system] First, a description will be given of the configuration of the information processing system 10. FIG.

[0022] The information processing system 10 has, for example, one or more information processing devices 1 (hereinafter also simply referred to as information processing devices 1) configured by one or more physical machines arranged on the cloud. The information processing system 10 also has, for example, an operation terminal (not shown) through which a developer D who generates (develops) a model structure uploads the model structure and data features (hereinafter also referred to as data features) suitable for the model structure, and an operation terminal (not shown) through which a user U who generates a learning model suited to his or her own business inputs necessary information.

[0023] The information processing device 1 may be, for example, one or more virtual machines or one or more containers deployed on a cloud. In addition, hereinafter, it is assumed that the learning model generated by the user U includes not only the model structure of the learning model, but also data features and parameters suitable for the model structure.

[0024] The information processing device 1 has, for example, a repository 1a (hereinafter also referred to as a memory unit 1a) that stores an ontology that defines model structures corresponding to multiple learning models and data features suitable for those model structures (hereinafter collectively referred to as model structures, etc.), and an inference engine 1b that infers model structures, etc. suitable for generating learning models.

[0025] Specifically, when developer D generates a new model structure, for example, developer D uploads the source code of the generated new model structure to information processing system 10 (information processing device 1). In this case, information processing device 1 adds a class corresponding to the uploaded new model structure to repository 1a in accordance with the ontology stored in repository 1a.

[0026] Meanwhile, the user U inputs an inference request, for example, a model structure required for generating a learning model, to the information processing system 10 (information processing device 1). Specifically, the user U inputs information including the field of the learning model to be generated (hereinafter also referred to as a task) and the cost that can be spent on generating the learning model as the inference request. Then, the information processing device 1 (inference engine 1b) refers to the ontology stored in the repository 1a, infers a model structure that matches the input inference request, and transmits information indicating the inferred model structure to the user U (for example, the operation terminal of the user U).

[0027] As a result, even if the user U does not have sufficient knowledge about AI, the user U can generate a learning model that meets his / her requirements by using the model structure inferred by the information processing device 1. Furthermore, the developer D can receive compensation from the user U by selling the model structure that he / she has generated to the user U, for example.

[0028] Although the following description will be given assuming that the information processing device 1 implements the inference engine 1b, the inference engine 1b may be implemented in another information processing device different from the information processing device 1.

[0029] [Configuration of information processing device] Next, a description will be given of the hardware configuration of the information processing device 1. FIG.

[0030] The information processing device 1 has the hardware configuration of a general-purpose computer device, and for example, as shown in Fig. 2, includes a CPU 101 which is a processor, a memory 102, a network interface 103, and a storage medium 104. Each part is connected to each other via a bus 105.

[0031] The storage medium 104 has a program storage area (not shown) for storing programs (not shown) for performing, for example, a process of storing an ontology in the repository 1a and a process of inferring a model structure or the like inferred from the ontology stored in the repository 1a (hereinafter, these processes are also collectively referred to as model inference processing). The storage medium 104 also has, for example, a storage area 110 for storing information used when performing the model inference processing.

[0032] The storage area 110 may function as the repository 1a, or a storage device (not shown) provided outside the information processing device 1 may function as the repository 1a. The storage medium 104 may be, for example, an HDD (Hard Disk Drive).

[0033] The CPU 101 functions as an inference engine 1b by executing a program loaded from the storage area 110 into the memory 102, and performs model inference processing.

[0034] The network interface 103 also communicates with, for example, a developer D's operation terminal and a user U's operation terminal.

[0035] [Model inference processing in the first embodiment] Next, a description will be given of the model inference processing in the first embodiment. Figures 3 and 4 are diagrams for explaining the model inference processing in the first embodiment.

[0036] 3, the CPU 101 of the information processing device 1 functions as a model receiving unit 11, an ontology description and instance generation unit 12, and a model storage unit 13 in cooperation with a program. Also, as shown in Fig. 4, the CPU 101 of the information processing device 1 functions as a request receiving unit 21, a rule generation unit 22, an inference execution unit 23, and a data output unit 24 in cooperation with a program. The inference execution unit 23 is a function corresponding to, for example, the inference engine 1b.

[0037] The model receiving unit 11 receives, for example, an upload of a new model structure created by a developer D.

[0038] The ontology description and instance generation unit 12 adds, for example, a class (subclass) corresponding to the new model structure accepted by the model acceptance unit 11 to the ontology stored in the repository 1a.

[0039] The model storage unit 13 stores the new model structure itself received by the model receiving unit 11 in the repository 1a, for example, in accordance with the ontology stored in the repository 1a.

[0040] The request receiving unit 21 receives, for example, an inference request for a learning model input by a user U to the information processing system 10.

[0041] The rule generation unit 22 generates rules (hereinafter also referred to as logical formulas) for inferring a model structure or the like corresponding to an inference request received by the request reception unit 21, for example, according to an ontology stored in the repository 1a. Below, a case where the rules are expressed in SWRL (Semantic Web Rule Language) will be described.

[0042] The inference execution unit 23 infers, for example, a model structure corresponding to the rule generated by the rule generation unit 22.

[0043] The data output unit 24 transmits, for example, the model structure inferred by the inference execution unit 23 to the user U (for example, the operation terminal of the user U).

[0044] [Details of Model Inference Processing in the First Embodiment] Next, details of the model inference process in the first embodiment will be explained below. Figures 5 to 16 are diagrams for explaining details of the model inference process in the first embodiment.

[0045] [What happens when a new model structure is uploaded] First, we will explain the model inference process, which is performed when developer D uploads a new model structure.

[0046] As shown in FIG. 5, the model receiving unit 11 receives, for example, an upload of a new model structure created by developer D (S11).

[0047] The ontology description and instance generation unit 12 adds, for example, a class corresponding to the new model structure received in the processing of S11 to the ontology stored in the repository 1a (S12).

[0048] The model storage unit 13 stores the new model structure received in the processing of S11 in the repository 1a, for example, in accordance with the ontology stored in the storage unit 1a (S13). A specific example of the ontology in this embodiment (the ontology stored in the repository 1a) will be described below.

[0049] [Specific example of ontology] 6 to 11 are diagrams illustrating specific examples of ontologies in this embodiment.

[0050] First, the top-level ontology will be explained. Figure 6 shows a specific example of the top-level ontology ONT1.

[0051] The ontology ONT1 shown in Figure 6 has the following classes: "DLG Ontology", "Goal (hereinafter also referred to as GoalExt)", "DLTask (hereinafter also referred to as TaskDomain)", "DLStructure (hereinafter also referred to as DLStructureExt)", "Data (hereinafter also referred to as DataExt)", and "Result".

[0052] Specifically, "Goal" defines, for example, each of the items that may be included in an inference request sent from user U. "DLTask" defines, for example, the field (business field) in which each learning model is developed. "DLStructure" defines, for example, each model structure developed by developer D. "Data" defines, for example, the data features of the learning data input to each model structure. "Result" defines, for example, each of the items that may be included in each inference result by inference engine 1b.

[0053] Furthermore, in the ontology ONT1 shown in Figure 6, "achive" is an attribute (property) indicating the relationship between "DLGOntology" and "Goal," indicating that "DLGOntology" achieves "Goal." "perform" is an attribute indicating the relationship between "DLGOntology" and "DLTask," indicating that "DLGOntology" performs "DLTask." "construct" is an attribute indicating the relationship between "DLGOntology" and "DLStructure," indicating that "DLGOntology" constructs "DLStructure." "learn" is an attribute indicating the relationship between "DLGOntology" and "Data," indicating that "DLGOntology" learns "Data." "generate" is an attribute indicating the relationship between "DLGOntology" and "Result," indicating that "DLGOntology" generates "Result." Explanation of other information included in Figure 6 is omitted.

[0054] Next, sub-level ontologies will be described. Figs. 7 to 11 are diagrams showing specific examples of sub-level ontologies. Specifically, Fig. 7 is a diagram explaining an ontology ONT2a for "Goal," Fig. 8 is a diagram explaining an ontology ONT2b for "DLTask," and Fig. 9 is a diagram explaining an ontology ONT2c for "DLStructure." Furthermore, Fig. 10 is a diagram explaining an ontology ONT2d for "Data," and Fig. 11 is a diagram explaining an ontology ONT2e for "DLResult."

[0055] First, we will explain the ontology ONT2a about "Goal".

[0056] As shown in FIG. 6, "Goal" has, for example, "Property" as a subclass. Furthermore, as shown in FIG. 7, "Property" has, for example, "FP (Functional Property)" and "NFP (Non-functional property)" as subclasses corresponding to an inference request received from a user U. For example, functional attributes related to each learning model are defined in "FP," and non-functional attributes such as price and performance are defined in "NFP." Specifically, attributes such as "Text Classification," "Language Translation," and "Image Segmentation" are defined in "FP." Furthermore, attributes related to QoS such as "Accuracy" and "Precision / Recall Rate" are defined in "NFP."

[0057] Next, we will explain the ontology ONT2b for "DLTask".

[0058] "DLTask" has subclasses such as "TextDomain," "ImageDomain," and "SignalDomain," each of which corresponds to a different data type, as shown in Fig. 8. Attributes such as "Classification," "Clustering," and "Ecoding" are defined for each subclass.

[0059] Next, we will explain the ontology ONT2c for "DLStructure".

[0060] As shown in FIG. 9, "DLStructure" has subclasses such as "SimpleDLStructure", "ComplexDLStructure", and "HypterParameter". "SimpleDLStructure" defines attributes such as "MPN (Multi-Perceptron Network)", "CNN (Convolutional Neural Network)", "RNN (Recurrent Neural Network)", and "Transformer", which define a basic model structure. "ComplexDLStructure" defines attributes such as "CNN-RNN", "TextPreTraining", which is a structure for pre-training text, "TextPreTraining", which is a structure for pre-training images, "Sequence2Sequence", and "Generative Adversarial Network (GAN)". "HypterParameter" defines attributes such as "LearningRate", "NoOfLayer", "NoOfInput", "NoOfOutput", "DropoutRate", and "NoOfEpoch".

[0061] Next, we will explain the ontology ONT2d for "Data."

[0062] As shown in Fig. 10, "Data" has subclasses such as "DataFormat" that defines the format of data and "DataFeature" that defines the features of data. "DataFeature" defines attributes such as "TFIDF" that indicates the frequency of words in TextData.

[0063] Next, the ontology ONT2e for "DLResult" will be explained.

[0064] As shown in FIG. 11, "DLResult" has classes such as "DLStructureResult" that defines the inference result for "DLStructure" and "DataResult" that defines the inference result for "Data".

[0065] [Processing performed when an inference request is entered] Next, the model inference process will be described, which is performed when a user U inputs an inference request for a learning model.

[0066] In this case, the information processing device 1 performs Goal Understanding of the content corresponding to the inference request sent from the user U, as shown in Fig. 12, prepares learning data (Data Preparation) as shown in Fig. 13, and further generates a learning model such as DL (Deep Learning Generation) as shown in Fig. 14. Details of the processing performed when the user U inputs an inference request for a learning model will be described below.

[0067] As shown in FIG. 15, the request receiving unit 21 receives, for example, an inference request for a learning model input by a user U to the information processing system 10 (S21).

[0068] Specifically, user U inputs an inference request to information processing device 1 for a classifier that can accurately classify documents with a close domain distance, such as documents about indoor sports and documents about outdoor sports, with an accuracy of 80% or more.

[0069] Furthermore, the request receiving unit 21 receives, for example, a plurality of pieces of learning data input by the user U to the information processing system 10 (S22).

[0070] Next, the rule generating unit 22 identifies, for example, an FP corresponding to the inference request received in the processing of S21 from the ontology stored in the repository 1a (S23).

[0071] Specifically, the inference request received in the process of S21 includes information to generate a "document classifier." Therefore, the rule generation unit 22 identifies, for example, "TextClassification," which is an attribute of "DLTask," as an FP.

[0072] Furthermore, the rule generating unit 22 identifies, for example, an NFP corresponding to the inference request received in the processing of S21 from the ontology stored in the repository 1a (S24).

[0073] Specifically, the inference request received in the process of S21 includes information to generate a "classifier for documents with close domain distances, such as documents about indoor sports and documents about outdoor sports," and a "classifier capable of accurate classification with an accuracy of 80(%) or more." Therefore, the rule generation unit 22 identifies, for example, "near," which is an attribute of "DomainDistance," which is a subclass of "NFP," "high," which is an attribute of "Accuracy," which is a subclass of "NFP," and "two," which is an attribute of "NoOfClass," which is a subclass of "NFP."

[0074] Next, the rule generating unit 22 identifies, for example, a task corresponding to the "FP" identified in the process of S23 (S25).

[0075] Specifically, the inference request received in the process of S21 includes information to generate a "document classifier." Therefore, the rule generation unit 22 identifies, for example, "TextClassification," which is an attribute of "DLTask."

[0076] Then, the rule generating unit 22 generates a rule (logical formula) indicating a model structure corresponding to the task identified in the process of S25 (S26). Specifically, the rule generating unit 22 generates the following rule (1).

[0077] task:TextClassfication(?task)^DLGT:DLStructure(?struct)^DLGT:implementedBy(?task,?struct)->result:DLStructureResult(?struct) · · · Rule (1)

[0078] In this case, "TextClassification" is changed to "TaskDomain" by the SUP() function, which returns "TaskDomain", the domain of "implementedBy", as shown in rule (2) below.

[0079] task:TextDomain(?task)^DLGT:DLStructure(?struct)^DLGT:implementedBy(?task,?struct)->result:DLStructureResult(?struct) · · · Rule (2)

[0080] In this way, the information processing device 1 can perform processing using the higher-level class "TaskDomain" regardless of the class corresponding to the task indicated by the inference request received in the processing of S21. Furthermore, even if there is a change in the classes lower than "TaskDomain" or if a new class is added, the information processing device 1 can similarly perform processing using the higher-level class "TaskDomain" thereof.

[0081] 16, the rule generating unit 22 generates a rule indicating the NFP identified in the process of S24 (S31). Specifically, the rule generating unit 22 generates the following rule (3).

[0082] property:NFP(?nfp)^property:DomainDistance(near)^property:Accuracy(high)^NoOfClass(two)^hasNFPVariableValue(?nfp,near)^hasNFPVariableValue(?nfp,high)^hasNFPVariableValue(?nfp,two)->result:PropertyResult(?nfp) · · · Rule (3)

[0083] Furthermore, rule generation unit 22 generates a rule indicating a model structure corresponding to the model structure indicated by the rule generated in the process of S26 and the "NFP" indicated by the rule generated in the process of S31 (S32). Specifically, rule generation unit 22 generates the following rule (4):

[0084] task:TaskDomain(?task)^DLGT:DLStructure(?struct)^DLGT:implementedBy(?task,?struct)^result:PropertyResult(?nfp)^DLGT:needSatisfy(?task,?nfp)->result:DLStructureResult(?struct) · · · Rule (4)

[0085] Then, rule generation unit 22 generates a rule indicating a data feature corresponding to the model structure indicated by the rule generated in the process of S32 (S33). Specifically, rule generation unit 22 generates the following rule (5).

[0086] result:DLStructureResult(?struct)^result:PropertyResult(?nfp)^DLGT:Data(?data)^DLGT:use(?struct,?data)^data:DataFeature(?df)^data:hasDataFeature(?data,?df)->result:DataFeatureResult(?df) · · · Rule (5)

[0087] In the process of S33, rule generation unit 22 may generate a rule indicating a parameter corresponding to the model structure indicated by the rule generated in the process of S32. Specifically, rule generation unit 22 may generate the following rule (6).

[0088] result:DLStructureResult(?struct)^HyperParameter(?hp)^hasHyperParameter(?struct,?hp)->result:HyperParameterResult(?hp) · · · Rule (6)

[0089] Next, the inference execution unit 23 infers a model structure and data features corresponding to the inference request received in the process of S21 from the rules generated in the processes of S32 and S33 (S34).

[0090] Specifically, the inference execution unit 23 specifies, for example, "CNN" as the model structure of the learning model and "tfidf-esa" as the data features of the learning data. In this case, the inference execution unit 23 may perform inference using an external service as necessary.

[0091] Furthermore, if the inference execution unit 23 has generated a rule indicating a parameter in the process of S33, it also infers a parameter corresponding to the inference request received in the process of S21 from the rule generated in the process of S33.

[0092] Then, the inference execution unit 23 converts the learning data received in the process of S22 so that it corresponds to the data features inferred in the process of S34 (S35).

[0093] Specifically, the inference execution unit 23 may search for an external service that can convert the learning data received in the process of S22, for example. Then, the inference execution unit 23 may convert the learning data received in the process of S22 by utilizing the searched external service.

[0094] Thereafter, the data output unit 24 outputs the model structure inferred in the process of S34 and the learning data converted in the process of S35 (S36).

[0095] Then, the user U generates a learning model corresponding to his / her own business, for example, by training the learning data converted in the process of S35 in the model structure inferred in the process of S34.

[0096] In the above example, we have explained the case where user U inputs an inference request for a document classifier, but the information processing device 1 can also perform inference on model structures, etc., even when user U inputs an inference request for a language translator, for example.

[0097] In this way, for example, when developer D generates a new model structure, the information processing device 1 in this embodiment adds the generated new model structure to the existing ontology as a subclass. Also, for example, when user U inputs an inference request for a learning model, the information processing device 1 generates rules corresponding to the input inference request based on the ontology, and performs inference on the model structure and data features according to the generated rules.

[0098] That is, the information processing device 1 in this embodiment generates an ontology by using semantic technology, and also infers model structures and data features by using rules independent of the ontology.

[0099] This enables the information processing device 1 in this embodiment to continue inferring model structures and data features even when the ontology is updated due to changes in existing model structures or the addition of new model structures.

[0100] Here, for example, if an information processing device 1 is constructed by using an R-DB instead of an ontology, the information processing device 1 will need to recreate the R-DB tables in response to changes in existing model structures or the addition of new model structures, which will increase the workload and costs required for maintaining the information processing device 1.

[0101] In this regard, the information processing device 1 according to the present embodiment performs model inference processing without using an R-DB, which makes it possible to reduce the workload and costs associated with changing an existing model structure or adding a new model structure. [Explanation of symbols]

[0102] 1: Information processing device 1a: Repository 1b: Inference Engine 11: Model reception desk 12: Ontology generation unit 14: Model memory section 21: Request reception unit 22: Rule generation unit 23: Inference execution unit 24: Data output section 101:CPU 102: Memory 103: Network interface 104:Storage medium 105: Bus

Claims

1. receiving an inference request for a model structure of a learning model required for generating a learning model and data features of learning data required for generating the learning model; a memory unit that stores an ontology having a hierarchical structure of classes, each of which corresponds to an item included in the inference request for each learning model, the first class having a first subclass corresponding to each item related to the function of each learning model and a second subclass corresponding to each item other than the function of each learning model, a second class corresponding to a task performed by each learning model, a third class corresponding to the model structure of each learning model, a fourth class corresponding to the data features of each learning model, and a fifth class corresponding to each item included in the inference result for each learning model; and generating a first logical formula indicating that the model structure indicated by the identified third instance and the data features indicated by the identified fifth instance are identified as at least a part of the inference result; outputting an inference result including the model structure and the data feature identified based on the generated first logical formula; A model inference program that causes a computer to execute processing.

2. In claim 1, In the process of generating the first logical formula, a fourth instance corresponding to the received inference request is identified from among the instances defined in the second subclass, and the first logical formula indicating that an instance corresponding to the identified second instance and fourth instance from among the instances defined in the third class is identified as the third instance is generated. A model inference program characterized by:

3. In claim 1, further comprising: receiving a plurality of pieces of training data to be used for generating the training model; Have the computer execute the process, In the process of generating the first logical formula, generating a second logical expression indicating that the received plurality of learning data should be converted so as to correspond to the data feature indicated by the identified fifth instance; A model inference program characterized by:

4. a request receiving unit that receives an inference request for a model structure of a learning model required for generating a learning model and data features of learning data required for generating the learning model; a rule generation unit that references a storage unit that stores an ontology having a hierarchical structure of classes, the first class corresponding to each item included in the inference request for each learning model, the first subclass corresponding to each item related to the function of each learning model, and a second subclass corresponding to each item other than the function of each learning model, a second class corresponding to a task performed by each learning model, a third class corresponding to the model structure of each learning model, a fourth class corresponding to the data features of each learning model, and a fifth class corresponding to each item included in the inference result for each learning model, and identifies a first instance corresponding to the received inference request from among the instances defined in the first subclass, identifies a second instance corresponding to the identified first instance from among the instances defined in the second class, identifies a third instance corresponding to the identified second instance from among the instances defined in the third class, and identifies a fifth instance corresponding to the identified third instance from among the instances defined in the fourth class, and generates a first logical formula indicating that the model structure indicated by the identified third instance and the data features indicated by the identified fifth instance are specified as at least a part of the inference result; a data output unit that outputs an inference result including the model structure and the data feature identified based on the generated first logical formula, 1. An information processing device comprising:

5. receiving an inference request for a model structure of a learning model required for generating a learning model and data features of learning data required for generating the learning model; a memory unit that stores an ontology having a hierarchical structure of classes, each of which corresponds to an item included in the inference request for each learning model, the first class having a first subclass corresponding to each item related to the function of each learning model and a second subclass corresponding to each item other than the function of each learning model, a second class corresponding to a task performed by each learning model, a third class corresponding to the model structure of each learning model, a fourth class corresponding to the data features of each learning model, and a fifth class corresponding to each item included in the inference result for each learning model; and generating a first logical formula indicating that the model structure indicated by the identified third instance and the data features indicated by the identified fifth instance are identified as at least a part of the inference result; outputting an inference result including the model structure and the data feature identified based on the generated first logical formula; A model inference method characterized by causing a computer to execute processing.

Images