Cloud server, edge server, and intelligent model generation method using the same

Abstract

To provide a method of generating and distributing an intelligence model using a complex computing environment comprising a cloud server and an edge server.SOLUTION: A method of generating an intelligence model is provided, the method comprising receiving, by an edge server, an intelligence model generation request from a user terminal, generating an intelligence model corresponding to the intelligence model generation request, and adjusting the generated intelligence model. Generating the intelligence model includes requesting, by the edge server, a cloud server to generate an intelligence model when failing to generate the intelligence model, and receiving an intelligence model generated by the cloud server.SELECTED DRAWING: Figure 1

Description

[[Technical Field]]

[0001] The present invention relates to machine learning-based intelligent model generation, distribution and management technologies.

[0002] Specifically, the present invention relates to a method for generating and distributing intelligent models performed on cloud servers and edge servers. [[Background Art]]

[0003] In order to effectively implement artificial intelligence services, it is necessary to easily secure and utilize high-quality intelligent models optimized for terminal requirements and application environments. Various conventional methods can be used to secure and utilize artificial intelligence models.

[0004] First, professionals may handle the entire process of developing an intelligent model. An artificial intelligence professional secures a dataset for training an artificial intelligence model, selects or designs and implements the structure of the artificial intelligence model, and then trains and tests the artificial intelligence model using the dataset until it operates at a desired performance level. The resulting intelligent model is deployed and utilized in the application environment.

[0005] In this case, since it is necessary to complete all the work including data securing, program development and training to obtain an intelligent model, it is difficult, costly and time-consuming to obtain an intelligent model. While this approach has the advantage of being able to obtain an intelligent model optimized for the application and application environment, variations in the performance and quality of the intelligent model may occur depending on the expertise of the person in charge of generating the intelligent model and the generation method.

[0006] Artificial intelligence professionals can also select and utilize suitable already published artificial intelligence models according to needs. They search for an appropriate artificial intelligence model through web search, obtain the model and related codes, and integrate them into application programs for utilization. Compared with the method of manually developing, training and obtaining a model, this method has lower difficulty, lower cost and can save time.

[0007] However, securing a model optimized for the application and its specific environment is difficult. Securing an optimized model requires collecting and building a large amount of training and evaluation data, which is costly and time-consuming. While information on the performance and quality of intelligent models can be referenced to ensure quality levels, the lack of such information makes it impossible to guarantee the performance and quality of the intelligent model.

[0008] Another way for AI experts or general developers to utilize AI services based on cloud platforms is to select the necessary AI services from those offered on the cloud platform, and then use the client-server programming interface provided by the service provider to request the execution of the AI ​​service and receive a response. Examples of such AI model services include Google Cloud, Microsoft Azure Cognitive Services, and Intel Watson.

[0009] This method is easy to implement and saves time and cost in acquiring intelligent models. However, because it relies on using pre-built intelligent models through a service, it is difficult to secure a model optimized for the user's specific application and environment. Furthermore, utilizing cloud platform services requires transmitting all user data to the cloud platform, which can lead to data security issues.

[0010] Another approach is to utilize the recently emerged cloud-based automated intelligent model generation services. Google's AutoML service generates and provides the optimal intelligent model desired by the user by training the model using training data provided by the user. This method is not only easy, inexpensive, and quick, but also makes it easier to secure a model optimized for the application and environment. Since the intelligent models are generated by specialized companies using pre-validated processes, the performance and quality are also good.

[0011] However, since it is necessary to build and provide a sufficient amount of dataset to train intelligent models, data acquisition is considerably difficult, costly, and time-consuming. Without data, it is impossible to acquire intelligent models. Furthermore, since all data must be transmitted to a cloud platform to train intelligent models, data security issues may arise. [Prior art documents] [Patent Documents]

[0012] [Patent Document 1] Korean Published Patent Publication No. 10-2020-0052449 (Title of Invention: Connected Data Architecture System for Artificial Intelligence Services and Control Method therefor) [Overview of the Initiative] [Problems that the invention aims to solve]

[0013] The object of the present invention is to provide a method for generating and distributing intelligent models through a hybrid computing environment consisting of cloud and edge computing.

[0014] Furthermore, an objective of the present invention is to secure an intelligent model that is optimized for application quickly and at low cost.

[0015] Furthermore, an objective of the present invention is to generate intelligent models optimized for application services and environments, even when there is no data or only a small amount of data is available.

[0016] Furthermore, an objective of the present invention is to prevent security and privacy infringement issues by dualizing the intelligent model generation process and preventing the external exposure of data. [Means for solving the problem]

[0017] An intelligent model generation method according to one embodiment of the present invention for achieving the above objective includes the steps of: an edge server receiving an intelligent model generation request from a user terminal; generating an intelligent model corresponding to the intelligent model generation request; and adjusting the generated intelligent model.

[0018] In this case, the step of generating the intelligent model may further include the step of requesting the cloud server to generate the intelligent model if the edge server fails to generate the intelligent model, and the step of receiving the intelligent model generated by the cloud server.

[0019] In this case, the cloud server may include a first cloud server and a second cloud server having a larger capacity than the first cloud server.

[0020] In this case, if the first cloud server fails to generate the intelligent model, it may request the second cloud server to generate the intelligent model.

[0021] In this case, the intelligent model generation request may include a task identifier, raw data, annotations, data disclosure scope, and target labels.

[0022] In this case, the step of generating the intelligent model may include the steps of selecting a basic intelligent model based on the intelligent model generation request, transforming the label list of the basic intelligent model to correspond to a target label list, and performing the transformed intelligent model learning.

[0023] In this case, the step of training the modified intelligent model may include a first training step using an already stored dataset and a second training step using the raw data included in the intelligent model generation request.

[0024] In this case, the step of requesting the cloud server to generate an artificial intelligence model may set raw data to be transmitted to the cloud server based on the data disclosure scope.

[0025] In this case, the step of adjusting the generated artificial intelligence model may be performed using raw data that has not been transmitted to the cloud server.

[0026] In addition, an edge server according to an embodiment of the present invention for achieving the above object may include: a communication unit that communicates with a user terminal and other servers; a storage unit that stores data for generating an artificial intelligence model; a model generation unit that generates an artificial intelligence model corresponding to an artificial intelligence model generation request; and an adjustment unit that adjusts the generated artificial intelligence model.

[0027] In this case, if the model generation unit fails to generate the artificial intelligence model, the communication unit may request a cloud server to generate the artificial intelligence model, and receive the artificial intelligence model generated by the cloud server.

[0028] In this case, the cloud server may include a first cloud server and a second cloud server having a larger capacity than the first cloud server.

[0029] In this case, if the first cloud server fails to generate the artificial intelligence model, the first cloud server may request the second cloud server to generate the artificial intelligence model.

[0030] In this case, the artificial intelligence model generation request may include a task identifier, raw data, annotations, a data disclosure scope, and target labels.

[0031] In this case, the model generation unit may select a basic artificial intelligence model based on the artificial intelligence model generation request, modify a label list of the basic artificial intelligence model to correspond to a target label list, and train the modified artificial intelligence model.

[0032] In this case, the communication unit may transmit the raw data to the cloud server based on the data disclosure scope.

[0033] In this case, the adjustment unit may adjust the intelligent model using raw data that has not been transmitted to the cloud server.

[0034] Furthermore, a cloud server according to one embodiment of the present invention for achieving the above objective includes a communication unit that receives intelligent model generation requests from edge servers, a storage unit that stores data for intelligent model generation, and a model generation unit that generates an intelligent model corresponding to the intelligent model generation request, wherein the intelligent model generation request may include a task identifier, raw data, annotations, data disclosure scope, and target labels.

[0035] In this case, if the model generation unit fails to generate the intelligent model, the communication unit may request the generation of the intelligent model from another cloud server.

[0036] In this case, the raw data of the intelligent model generation request may be transmitted by the edge server based on the data disclosure scope. [Effects of the Invention]

[0037] According to the present invention, a method for generating and distributing intelligent models through a hybrid computing environment consisting of cloud and edge computing can be provided.

[0038] Furthermore, the present invention makes it possible to quickly secure an intelligent model optimized for application at low cost.

[0039] Furthermore, the present invention can generate intelligent models optimized for application services and environments, even when there is no data or only a small amount of data is available.

[0040] Furthermore, this invention can prevent security and privacy infringement issues by dualizing the intelligent model generation process and preventing data from being exposed to the outside. [Brief explanation of the drawing]

[0041] [Figure 1] This is a flowchart showing an intelligent model generation method according to one embodiment of the present invention. [Figure 2] This flowchart provides a more detailed explanation of the intelligent model generation method according to one embodiment of the present invention. [Figure 3] This figure shows the configuration of an intelligent model distribution system according to one embodiment of the present invention. [Figure 4] This is an example of an intelligence requirements profile that requests an intelligent model for image classification. [Figure 5] This is an example of an intelligence requirements profile that requests an intelligent model for object detection tasks. [Figure 6] This is an example of an intelligence requirement profile that requests an intelligent model for semantic-based image segmentation. [Figure 7] This is a block diagram showing an intelligent repository structure according to an embodiment of the present invention. [Figure 8] This figure shows an example of a dataset for an intelligent model generation method according to one embodiment of the present invention. [Figure 9] This is a diagram conceptually illustrating the structure of AlexNet. [Figure 10] This is a flowchart showing the intelligent model distribution process of the present invention. [Figure 11] This flowchart shows the intelligent model generation process according to an embodiment of the present invention. [Figure 12] This is an example of generating a list of standard correct answer labels. [Figure 13] This is an example of how an edge server can be modified based on the data exposure scope of the intelligence requirements profile shown in Figure 4. [Figure 14] This is a block diagram showing the structure of an edge server according to one embodiment of the present invention. [Figure 15]A diagram showing the structure of a cloud server according to one embodiment of the present invention. [Figure 16] This diagram shows the configuration of the computer system according to the embodiment. [Modes for carrying out the invention]

[0042] The advantages and features of the present invention, as well as methods for achieving them, will become apparent with reference to the embodiments described below in detail with the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below and can be embodied in a variety of different forms, provided only to complete the disclosure of the present invention and to fully inform those who are ordinaryly skilled in the art to which the present invention pertains, and the present invention is defined only by the scope of the claims. Throughout the specification, the same reference numerals refer to the same components.

[0043] Even if terms such as "first" or "second" are used to describe various components, such components are not limited by such terminology. Such terms may simply be used to distinguish one component from another. Therefore, the first component referred to below may be the second component within the technical concept of the present invention.

[0044] The terms used herein are for illustrative purposes only and are not intended to limit the invention. In this specification, the singular form includes the plural form unless otherwise specified in the text. As used herein, “comprises” or “comprising” implies that the components or steps mentioned do not exclude the presence or addition of one or more other components or steps.

[0045] Unless otherwise defined, all terms used herein may be interpreted in a way that is commonly understood by a person of ordinary skill in the art to which the present invention pertains. Furthermore, terms defined in commonly used dictionaries shall not be interpreted ideally or excessively unless explicitly defined otherwise.

[0046] Hereinafter, embodiments of the present invention will be described in detail with reference to the attached drawings. When describing with reference to the drawings, the same or corresponding components will be denoted by the same reference numerals, and redundant descriptions thereof will be omitted.

[0047] Figure 1 is a flowchart showing an intelligent model generation method according to one embodiment of the present invention.

[0048] The method for generating and distributing intelligent models according to one embodiment of the present invention can be performed on edge servers and cloud servers. However, intelligent model generation can also be performed only on edge servers in response to intelligent model generation requests from user terminals, and the scope of the present invention is not limited thereto.

[0049] Referring to Figure 1, the method according to an embodiment of the present invention includes the steps of: an edge server receiving an intelligent model generation request from a user terminal (S110); generating an intelligent model corresponding to the intelligent model generation request (S120); and adjusting the generated intelligent model (S130).

[0050] In this case, the step of generating the intelligent model (S120) may further include, if the edge server fails to generate the intelligent model, the step of requesting the cloud server to generate the intelligent model and the step of receiving the intelligent model generated by the cloud server.

[0051] In this case, the cloud server may include the first cloud server and a second cloud server having a larger capacity than the first cloud server.

[0052] In this case, if the first cloud server fails to generate the intelligent model, it may request the second cloud server to generate the intelligent model.

[0053] In this case, the intelligent model generation request may include a task identifier, raw data, annotations, data disclosure scope, and target labels.

[0054] In this case, the step of generating the intelligent model (S120) may include the steps of selecting a basic intelligent model based on the intelligent model generation request, transforming the label list of the basic intelligent model to correspond to the target label list, and training the transformed intelligent model.

[0055] In this case, the step of training the modified intelligent model may include a first training step using an already stored dataset and a second training step using the raw data included in the intelligent model generation request.

[0056] In this case, the step of requesting the cloud server to generate an intelligent model may involve setting the raw data to be transmitted to the cloud server based on the data disclosure scope.

[0057] In this case, the step of adjusting the generated intelligent model (S130) may be performed using raw data that has not been transmitted to the cloud server.

[0058] Figure 2 is a flowchart illustrating in more detail the intelligent model generation method according to one embodiment of the present invention.

[0059] Referring to Figure 2, the intelligent model generation method according to one embodiment of the present invention may be performed on a user terminal (10), an edge server (20), and a cloud server (30).

[0060] The user terminal (10) requests the edge server (20) to generate an intelligent model necessary for providing the service (S11). Upon receiving the intelligent model generation request, the edge server (20) determines whether it can generate the intelligent model within the edge server (20) (S12). If it is possible to generate the intelligent model (S12), the edge server generates the intelligent model (S13), fine-tunes the intelligent model (S20), and transmits it to the user terminal (10) (S21). In this case, if the intelligent model is generated within the edge server, the step of fine-tuning the intelligent model (S20) may be omitted.

[0061] If the edge server (10) is unable to generate an intelligent model (S120), it transmits a request for intelligent model generation to the cloud server (30) (S14). In this case, the cloud server (30) refers to a server having larger computing resources than the edge server, and the scope of the present invention is not limited by this terminology.

[0062] Upon receiving the intelligent model generation request, the cloud server (30) determines whether or not intelligent model generation is possible (S15). If intelligent model generation is possible, it generates the intelligent model and transmits it to the edge server (20). The edge server fine-tunes the received intelligent model (S20) and transmits it to the user terminal (10) (S21).

[0063] If it is not possible to generate an intelligent model on the cloud server (S15), a request is made to another cloud server to generate the intelligent model (S17). In this case, the other cloud server corresponds to a server with larger computing resources than the cloud server (30). The cloud server (30), having received the intelligent model from the other cloud server (S18), transmits it to the edge server (S19). The edge server fine-tunes the received intelligent model (S20) and transmits it to the user terminal (10) (S21).

[0064] In this case, the step of requesting the generation of an intelligent model from the other cloud server (S17) may be repeated or performed hierarchically until a cloud server capable of generating the intelligent model is found. The present invention will be described in detail below through detailed embodiments.

[0065] Figure 3 shows the configuration of an intelligent model distribution system according to one embodiment of the present invention.

[0066] Referring to Figure 3, the system according to an embodiment of the present invention may consist of a terminal (100), an edge server (150), and a cloud server (200).

[0067] The terminal (100) is a device that provides intelligent services such as robots and smart speakers, and requests and utilizes intelligent models.

[0068] The edge server (150) is a computing system connected to a terminal via a network, and is generally located in a location physically close to the terminal. For example, if the terminal is a cafeteria service robot, the edge server (150) may be a server computer located within the cafeteria where the robot is operated.

[0069] Unlike cloud servers, edge servers (150) can be used in the region where the intelligent model is utilized, for example, within a store such as a restaurant, and may be managed by the operator of that region. In this case, the edge server (150) can solve data security problems by separating the data that should be provided for the generation and utilization of the intelligent model in accordance with the provisions of the Personal Information Protection Act or rules determined by the operator, into data that can be leaked externally and data that cannot, and transmitting only the data that can be leaked externally to an external server.

[0070] According to the intelligent model generation method of the present invention, an intelligent model generated in the cloud using publicly available data can be retrained and optimized as security data on an edge server, thereby securing the desired intelligent model while resolving data security issues.

[0071] A cloud server (200) is a computing device operated remotely, possessing more computing resources than terminals or edge servers, and capable of processing requests from a large number of edge servers or terminals. According to the present invention, cloud servers can be hierarchically linked across multiple stages. If it is not possible to generate and distribute intelligent models on a cloud server directly linked to an edge server, the request is transmitted to the next stage cloud server for processing. Preferably, the further away the cloud server is, i.e., the higher the stage, the larger the storage capacity and computing resources, allowing it to store more intelligent models and datasets, and enabling the generation and distribution of a larger number of intelligent models.

[0072] The edge server (150) and cloud server (200) perform the function of generating and distributing intelligent models through the intelligent repository (203), intelligent repository interface (204), and intelligent administrator (201).

[0073] The intelligent repository (203) stores and manages the information necessary to generate and optimize intelligent models. This information includes labels indicating the objects handled by the intelligent model, datasets used to train and evaluate the intelligent model, the structure and content of the intelligent model, and programs that perform inference, training, transfer learning, etc., based on the intelligent model.

[0074] The intelligent repository interface (204) is a message or programming interface used to store and access all of the aforementioned information.

[0075] The intelligence administrator (201) performs the function of generating and distributing intelligence requested from terminals, edge servers, or lower-level cloud servers using the intelligence repository (203). In this case, if the server cannot generate an intelligence model independently, it can transmit the intelligence request to the higher-level server and have the intelligent model distributed.

[0076] The intelligence requirements profile (110) is a data structure that records the specifications of the intelligence required by the terminal, and includes the functions that the intelligent model should perform and the data necessary for training the intelligent model.

[0077] In one embodiment of the present invention, intelligent model transmission can be implemented by transmitting together the generated "intelligent model" and a "program" that can drive the model and execute its functions. The terminal (100) can input the transmitted "intelligent model" and execute the "program" to utilize the inference function of the generated intelligent model.

[0078] The structure of the intelligence requirements profile (110) according to an embodiment of the present invention will be described in detail below.

[0079] The generation and distribution of intelligent models are carried out by transmitting an intelligent requirements profile (110). The intelligent requirements profile (110) contains cue information necessary for generating an intelligent model. In one embodiment of the present invention, the intelligent requirements profile (110) includes task details and objective data.

[0080] The task details describe the tasks that the intelligent model should perform and are used to search for and select intelligent models on edge servers and cloud servers. In one embodiment of the present invention, the task details include a task identifier, input information, and output information.

[0081] A task identifier is an item that indicates the work performed by the intelligent model, and its values ​​may include, for example, classification, detection, semantic segmentation, instance segmentation, natural language translation, and image captioning.

[0082] The input information describes the format and content of the data provided as input to the intelligent model. In one embodiment, the input information can be described based on modalities such as image, video, audio, and text.

[0083] The output information describes the format and content of the output data that the intelligent model outputs after receiving and processing the input. In one embodiment, the output information can be described as a class identifier (Class ID), bounding box, pixel-wise image mask, etc. Table 1 below shows some examples of task details. [Table 1]

[0084] The target data (110-2) is data that can be used to train or optimize an intelligent model, and includes raw data in various forms such as video, audio, and text (110-3), data annotations (110-4) indicating the information that the intelligent model should output based on the input of the raw data, the scope of publication for each data (110-5), and target correct answer labels (110-6) that indicate the subject that the intelligent model should handle, even if raw data is not provided.

[0085] Data annotations (110-4) show the correct answers for each item in the raw data (110-3). The correct answers may differ in form depending on the type of task performed by the intelligent model, such as classification, detection, or segmentation.

[0086] The disclosure scope (110-5) indicates the extent to which each raw data and data annotation can be made public. By restricting the data disclosure scope, the objective of protecting the private information of individuals and companies utilizing intelligent models can be achieved. In one embodiment, the disclosure scope can be described as "Regional" and "All-Area". When an edge server receives an intelligent request profile, it transmits data marked "All-Area" to the cloud server, and protects data marked "Regional" by processing it independently without transmitting it to the cloud server.

[0087] The objective correct answer label list (110-6) contains the names of the objects that the intelligent model detects or recognizes. If there is no data available to train the intelligent model, create this list and include it in the profile.

[0088] Figure 4 shows an example of an intelligence requirement profile that demands an intelligent model for image classification.

[0089] Referring to Figure 4, it can be seen that an intelligent model is requested that receives image input through the task details, classifies it, and outputs a class identifier. The target data includes raw image files that can be used for training, and the correct classification answers for each image file are included in the data annotations. It can also be seen that the scope of disclosure for data security is included.

[0090] Figure 5 shows the intelligence requirements profile that requests an intelligent model for object detection tasks.

[0091] Referring to Figure 5, we can see that the task details request an intelligent model that receives image input, detects objects, and assigns class identifiers to the detected regions. The target data includes raw image files that can be used for training, with data annotations containing the regions and correct class answers for the objects contained in each image.

[0092] Figure 6 shows the intelligence requirements profile that requests an intelligent model for semantic-based video segmentation.

[0093] Referring to Figure 6, it can be seen that the task details request an intelligent model that receives image input, divides the object's region into an image mask, and assigns class identifiers to the divided regions. The target data includes raw image files that can be used for training, and the data annotations include the names and class identifiers of the images to be used as image masks. As shown in Figures 4 to 6, intelligent requirement profiles can be configured and used to suit various tasks.

[0094] Figure 7 is a block diagram showing an intelligent repository structure according to an embodiment of the present invention.

[0095] Referring to Figure 7, the intelligent repository (203) includes a label dictionary (300), a dataset storage (400), an intelligent model storage (500), an intelligent model typology dictionary (600), and an intelligent model utilization code dictionary (700).

[0096] The label dictionary (300) is a dictionary for converting labels that have the same meaning but are represented by different strings or numbers into a standard vocabulary. The standard vocabulary is represented by a label identifier (301-1) that represents each label.

[0097] For example, a globally unique identifier such as a UUID can be used. The contents of the label dictionary in one embodiment are shown in [Table 2] below. [Table 2]

[0098] The label dictionary contains a list of label items, and each label item includes a label identifier and a natural language label. According to the dictionary in [Table 2], "cat", "猫", and "Chat" are all converted into the standard vocabulary "L0000001". The label dictionary can be constructed based on a dictionary database such as WordNet, as in the case of ImageNet, and can be extended to various languages through a translation machine. Methods for intelligently constructing and managing a label dictionary are not included in the scope of the present invention.

[0099] The dataset storage (400) stores a dataset (401) used for training and evaluating an intelligent model, as well as raw data items (402) and correct answer data items (403) that constitute the dataset.

[0100] In one embodiment of the present invention, the dataset is composed of a dataset identifier (401-1) and a correct answer data list (401-2). The dataset identifier (401-1) is a proper name that can uniquely distinguish and indicate the dataset. The correct answer data list (401-2) is a list of correct answer data identifiers (403-1) that point to correct answer data items (403). By referring to this list, all raw data items (402) and correct answer data items (403) constituting the dataset can be viewed.

[0101] The raw data item (402) includes a raw data identifier (402-1) that uniquely identifies the item, raw data (402-2) which is original data used for training and evaluating an intelligent model, and a raw data type (402-3) that describes the format of raw data such as images, videos, and audios.

[0102] The correct answer data item (403) includes a correct answer data identifier (403-1) that uniquely identifies the item, a raw data identifier (402-1) that points to the raw data to which the correct answer applies, correct answer data (403-2) that describes a correct answer label identifier, and task details (403-3) for which the correct answer can be used.

[0103] Figure 8 shows an example of a dataset for an intelligent model generation method according to one embodiment of the present invention.

[0104] Referring to Figure 8, the correct answer data item A0100111 designates L0001010 (airplane) as the correct answer label for the photograph indicated by the raw data item RD1340101. As can be seen from the examples of A0100111 and A0100133, multiple correct answer data items can refer to a single label (L0001010). The reverse is also true, as a single raw data item can be assigned multiple correct answer labels. A single raw data item can also be assigned multiple correct answers for different tasks. For example, a single photograph can be assigned a correct answer for classification, a correct answer for detection, and a correct answer for segmentation.

[0105] In Figure 8, A01000116 and A0100117 each provide different correct answers to a single raw data (RD1387478). A01000116 is a correct answer that designates "face (L1034962)" as the correct answer label for a classification task in a photograph containing a face, while A0100117 is a correct answer for a detection task that detects the facial region in a photograph and classifies it as "face". In the embodiment of the present invention, a single dataset is constructed by describing a list of correct answer data items. "Dataset 1" in Figure 8 is a dataset that can be used to train and evaluate an intelligent model that receives an image input and classifies it into one of airplane, automobile, ostrich, or face. "Dataset 2" is a dataset that can be used to train and evaluate an intelligent model that detects faces from images.

[0106] The intelligent model storage (500) stores numerous intelligent models (501). Each intelligent model (501) consists of a pair of intelligent model data (502) and intelligent model metadata (503).

[0107] Intelligent model data (502) is data necessary to run the intelligent model. In one embodiment of the present invention, the intelligent model data (502) consists of an intelligent model identifier (502-1) that uniquely identifies the model, a model type identifier (502-2) that can be used to understand the structure of the model, model parameter values ​​(502-3), and task details (502-4).

[0108] An intelligent model identifier (502-1) is an ID that uniquely distinguishes an intelligent model globally and can be specified using a global identifier such as a UUID.

[0109] The model type identifier (502-2) is a value that refers to the intelligent model type (601) which describes the structural details of the model. For example, the model type of an intelligent model based on an artificial neural network refers to neural network structure information that describes how neurons and layers are configured and connected.

[0110] Model parameter values ​​(502-3) are the actual values ​​of the various parameters that make up the model. In the case of a neural network model, this includes values ​​such as weights and biases. Since there are various methods for describing the model structure and parameter values ​​of intelligent models based on neural networks and machine learning, these methods should be utilized in intelligent model data technology. For example, ONNX (Open Neural Network Exchange) is a representative industry standard for describing model structure and parameter values.

[0111] Task details (502-4) are information describing what tasks the intelligent model will perform, and are identical to task details (110-1) included in the intelligence requirements profile (110). By comparing task details (110-1) described in the intelligence requirements profile with task details (502-4) included in the intelligent model data (502), an intelligent model that can appropriately perform the required functions can be selected.

[0112] The intelligent model metadata (503) includes descriptive information such as the method of generating the intelligent model, its functions, and its quality. The intelligent model metadata can be used not only as a reference for selecting and utilizing intelligent models, but also as a clue to determine the similarity between intelligent models and to identify intelligent models with quality problems. In one embodiment of the present invention, the intelligent model metadata (503) consists of a dataset identifier (503-1), a list of correct answer labels (503-2), a base model (503-3), a training history (503-4), performance evaluation information (503-5), and a quality history (503-6).

[0113] The dataset identifier (503-1) indicates the dataset used to train the intelligent model. It has one dataset identifier (401-1) as its value from among the datasets (401) stored in the dataset storage (400).

[0114] The correct answer label list (503-2) describes the correspondence between the output values ​​of the intelligent model and the label identifiers (302). When the intelligent model outputs a class ID, this value is the class index. For example, if the intelligent model performs a task of classifying images into two classes, dogs and cats, the output value of the intelligent model will be either 0 or 1. Assuming that 0 represents a cat and 1 represents a dog, the correct answer label list (503-2) describes this correspondence. The correspondence is described by specifying the label identifier (301-1) for each class ID. Based on the label dictionary in [Table 2], if the correct answer label list (503-2) is written as {0:L0000001, 1:L0000002}, then if the output of the intelligent model is 0, it can be interpreted as meaning "cat" with label identifier L0000001, and if it is 1, it can be interpreted as meaning "dog" with label identifier L0000002.

[0115] The base model (503-3) is the unique ID of the model used to train the intelligent model. For example, if the model M was trained through fine-tuning or other transfer learning methods based on model M, the intelligent model identifier (502-1) of M should be entered in the base model field. Leave this field blank if there is no base model.

[0116] The training history (503-4) includes parameter values ​​related to training the intelligent model and data generated during the process. For example, it may include not only the learning rate, batch size, and initial weights of the neural network, but also all information on what data was input in each training epoch, how the weights changed, how parameter values ​​that adjust the training process, such as the learning rate, were changed, and how the loss value changed.

[0117] Performance evaluation information (503-5) is data describing the performance of an intelligent model, including the dataset used for evaluation and performance values. It describes the dataset used for evaluation or the unique IDs of the evaluation data items, performance values ​​according to the model's evaluation scale, and the evaluation environment. For example, in the case of an image classification model, it can describe the unique IDs of all image data used for performance evaluation, performance values ​​such as classification accuracy and execution speed per image (fps), and the specifications of the CPU, GPU, and RAM of the system used for evaluation. Since performance values ​​may differ depending on the data configuration and evaluation environment used for performance evaluation, evaluation information should be continuously added to the performance evaluation information when the data and environment differ.

[0118] The quality history (503-6) includes various problem data that occurred during the utilization of the intelligent model. For example, the items in the quality history may include a unique problem number, problem description information, and problem severity information. Problem severity can be described in stages such as "serious," "moderate," and "negligible." Each quality history item can be made more reliable by including information such as the user who provided the quality history information, usage time, and self-assessed performance during use. The quality history information can be stored in a separate quality history storage so that it can be shared and tracked among users utilizing various intelligent models. The quality history of the intelligent model can be referenced within the intelligent model metadata by including the unique number of the information item stored in the quality history storage.

[0119] The Intelligent Model Type Dictionary (600) stores Intelligent Model Types (601), which are information structures that formally describe the structure of various intelligent models. An Intelligent Model Type (601) includes a Model Type Identifier (601-1) used to uniquely distinguish the model type, a Model Type Structure Detail (601-2) that formally describes the structure of the model, and a Task Detail (601-3) that describes the tasks that the model type can perform.

[0120] The model type structure details (601-1) are an information structure that formally describes an intelligent model, and must be read through a program to generate an intelligent model, which can then be trained and tested. In one embodiment of the present invention, ONNX (Open Neural Network Exchange), which describes the intelligent model on a deep learning platform as a computational graph structure, can be utilized. This method involves converting the structure of the intelligent model into ONNX format, storing it in the model type structure details (601-2), and then utilizing it. After selecting the required intelligent model type using the model type identifier (601-1), the model type structure details (601-2) can be loaded, and then the intelligent model can be trained or tested. It can also be used to determine whether the structures of different intelligent models are identical or similar to each other.

[0121] The task identifier (601-3) is the same information as the task identifier included in the task details (502-4) of the intelligent model (501), and describes the tasks that the intelligent model (501) trained based on the model type (601) can perform. For example, if the structure details (606-2) of the model type (601) is an AlexNet structure, it can perform classification tasks; if it is an R-CNN structure, it can perform detection tasks; and if it is a U-Net structure, it can perform segmentation tasks.

[0122] Figure 9 is a conceptual diagram illustrating the AlexNet structure.

[0123] Table 3 below shows the AlexNet structure from Figure 9 converted to ONNX details. [Table 3] TIFF0007927465000004.tif76164

[0124] The method according to an embodiment of the present invention can convert the deep learning model shown in Figure 9 into ONNX details and store them in an intelligent model type dictionary. The stored intelligent model type structure details can then be restored to a deep learning framework model such as PyTorch through a reconstruction process and used for training and testing.

[0125] The intelligent model utilization code storage (700) stores intelligent model utilization codes (701), which are programs that perform various tasks on intelligent models. An intelligent model utilization code (701) consists of a code identifier (701-1) used to uniquely identify the code, a code type (701-2) that describes the task the code performs, an execution code (701-3), and a compatibility model (701-4) that records the intelligent models that the code can handle.

[0126] In one embodiment of the present invention, the values ​​of code types (701-2) can be described as inference, training, fine-tuning, knowledge distillation, compression, etc. An inference code performs the function of loading an intelligent model (501), receiving input data, and providing output values ​​calculated through the intelligent model. A training code is an intelligent model type (601) that performs the function of generating an initial intelligent model and then training the intelligent model using a dataset (401) or an intelligence requirement profile (210). A fine-tuning code performs the function of loading an intelligent model (501), modifying the intelligent model structure according to the target correct answer labels (110-2) described in the intelligence requirement profile (210), and then training the intelligent model based on a dataset (401) or target data (110-1).

[0127] In one embodiment of the present invention, the code (701-3) utilizes a container runtime such as docker, containerd, or CRI-O to overcome compatibility issues due to different operating environments and to standardize the execution method. For example, a docker container containing code for training an AlexNet model, installed with a Linux OS, CUDA toolkit, Python, PyTorch framework, etc., can be stored and utilized within the intelligent model utilization code (701) code (701-3). Furthermore, a command script capable of driving the container can also be stored and utilized together with the code (701-3).

[0128] Figure 10 is a flowchart showing the intelligent model distribution process of the present invention.

[0129] Referring to Figure 10, the terminal generates an intelligent request profile (110) that includes the tasks to be performed by the intelligent model and the training data, and requests the edge server to generate and distribute the intelligent model (S1000).

[0130] In other words, step (S1000) is a step in which the terminal requests an intelligent model necessary for providing intelligent services. Those involved in providing intelligent services, such as the terminal manufacturer, installation specialist, and user, generate an intelligent request profile (110) through any user interface (which can be provided by the terminal, edge server, or cloud server). The user interface can be provided in various forms, such as a web interface, a graphics user interface, a chatbot, or a command window.

[0131] The intelligent administrator of the terminal transmits an intelligent request profile (110) to the edge server (150) to request the distribution of an intelligent model. The structure and contents of the intelligent request profile are shown in the examples in Figures 3 to 5.

[0132] The intelligent administrator (201) of the edge server selects a "baseline intelligent model" based on the task details and data contained in the intelligent request profile (110) received while referring to the intelligent repository (203) via the intelligent repository interface (204), and generates a new intelligent model (S1001) by constructing a "training / evaluation dataset" and training it. If the generation of the intelligent model is successful, the generated intelligent model and dataset information are added to the intelligent repository (203). The intelligent administrator (201) transmits the generated intelligent model and the intelligent model driving program to the terminal, and the terminal utilizes the intelligent model.

[0133] If the intelligent administrator (201) of the edge server fails to generate an intelligent model (S1001), the intelligent administrator (201) modifies the intelligent request profile (110) according to rules determined for purposes such as data security, and then transmits the intelligent request profile to the cloud server to request the generation and distribution of an intelligent model (S1002).

[0134] The intelligent administrator (201) of the cloud server attempts to generate an intelligent model using the same method as the intelligent administrator (201) of the edge server (S1003). If the generation of the intelligent model is successful, the generated intelligent model and related datasets are registered in the intelligent repository (203). If the generation of the intelligent model fails, an intelligent request profile (110) is transmitted to the next cloud server to request the generation and distribution of an intelligent model (S1002).

[0135] When the intelligent administrator (201) of the cloud server successfully generates an intelligent model, it transmits the intelligent model and the intelligent model driving program to the edge server that requested distribution (S1004).

[0136] If the edge server has separately stored data during the process of modifying the intelligence request profile (110), it optimizes the intelligent model using that data. The optimized intelligent model is added to the edge server's intelligence repository (203) and transmitted to the terminal (S1005). The terminal utilizes the distributed intelligent model and intelligent model driving program (S1006).

[0137] Steps S1001 and S1003 in Figure 10 include the process of generating an intelligent model based on the intelligence requirements profile (110). Below, the process of generating an intelligent model that performs a classification task will be described in detail according to one embodiment of the present invention.

[0138] Figure 11 is a flowchart showing the intelligent model generation process according to an embodiment of the present invention.

[0139] Referring to Figure 11, the intelligence manager (201) searches the intelligence repository (203) based on the task details (110-1) and objective data (110-2) to select a compatible intelligence model M (S2001) in order to generate an intelligence model capable of performing the tasks described in the intelligence requirement profile (110).

[0140] In one embodiment of the present invention, the selection of a compatible intelligent model follows a method of comparing the task details (110-4) described in the intelligent requirement profile (110) with the task details (502-4) of the intelligent model (501) stored in the intelligent model storage (500), and selecting the intelligent model if they are identical. This is called the primary selection process.

[0141] If two or more intelligent models are selected in the initial screening, a secondary screening process is performed based on the target data (110-2) of the intelligence requirement profile (110). In one embodiment of the present invention, this process is performed by calculating the similarity between the standard target label list and the correct answer label list (2000-2) of the compatible intelligent models.

[0142] The standard objective label list is the result of combining the correct answers in the objective data (110-1) and the correct answer labels included in the objective correct answer label list (110-2), and then converting each label to standard vocabulary through the lexical analyzer (202). The lexical analyzer (202) refers to the label dictionary (300) for lexical conversion. If lexical analysis fails, it is considered that the generation of the intelligent model has failed.

[0143] Figure 12 shows an example of generating a list of standard correct answer labels.

[0144] The similarity between the standard objective label list and the correct answer label list (2000-2) for compatible intelligent models can be calculated in various ways. In one embodiment of the present invention, the similarity between two label lists can be calculated through the Jaccard Index. The higher the similarity between the two label lists, the higher the priority for selecting the intelligent model.

[0145] After the secondary selection process, if there are two or more intelligent models with the same priority, a tertiary selection process is performed to select the model with superior performance and no quality issues. In one embodiment of the present invention, the tertiary selection process is performed based on the intelligent model metadata (503) of the intelligent repository (203). As described in the example below, general performance and quality indices can be referenced, or a method for measuring performance against target data can be utilized.

[0146] The method for referencing general performance involves comparing the performance evaluation information (503-5) explicitly stated in the intelligent model metadata (503) and selecting the best-performing intelligent model.

[0147] The method for referencing the quality index involves comparing the quality history (503-6) information explicitly stated in the intelligent model metadata (503) to select intelligent models that are free from quality issues.

[0148] The method for measuring performance on target data involves evaluating the performance of intelligent models on target data (110-2) included in the intelligence requirements profile and selecting the intelligent model with the highest performance. To this end, an intelligent model utilization code (701) capable of performing inference work on the model type (601) of the intelligent model to be evaluated is selected in the intelligent model utilization code storage (700). Then, the inference function can be performed on the target data (110-1) through the candidate intelligent model (501) to evaluate its performance.

[0149] The performance of the target data, general performance, and quality index can be weighted differently depending on the usage of intelligence and the environment, and adjusted to select a more appropriate intelligence model.

[0150] Through the above three-stage selection process, the intelligent model M with the highest priority is selected. Once M is selected, utilization codes that perform various functions on M are viewed and secured. Specifically, in the intelligent model utilization code storage (700), intelligent model utilization codes (701) that include the model type (502-2) of M in the compatible models (701-4) are selected and the codes are secured. In one embodiment of the present invention, each code consists of a container and a container-driven script pair, and each code can be used to perform functions such as inference, training, fine-tuning, and knowledge distillation. That is, after selecting the intelligent model M, codes such as code C1 that receives input from M and performs inference, code C2 that performs training, and code C3 that performs fine-tuning are secured.

[0151] Next, if the list of correct answer labels that M can process does not exactly match the standard target list of correct answer labels, the structure of M is modified to generate M1 (S2002).

[0152] For example, if M is an intelligent model that classifies images into 100 classes, and the standard objective label list contains only 10 correct answers, then the objective of this step is to optimize M so that it can classify only the 10 classes included in the standard objective label list.

[0153] If M is a Convolutional Neural Network structure for classification, then the final classification hierarchy should have 100 nodes and be connected to the convolution hierarchy in a fully connected structure. In this step, the classification hierarchy of M containing 100 output nodes is removed, and instead, a classification hierarchy containing 10 output nodes is generated and connected.

[0154] In one embodiment of the present invention, when generating M1, one additional output node can be added beyond the standard target correct label count. The added node is an "unknown" node, and by training it to activate when data that does not correspond to the standard target correct label is input to the intelligent model, the probability of false positives can be reduced and the classification accuracy can be improved.

[0155] This step (S2002) is not necessary if the length of the correct answer label list and the standard objective label list for M are the same.

[0156] Next, a dataset D for training M1 is constructed (S2003). D is constructed by browsing the dataset storage (400) of the intelligent repository (203) based on the standard objective label list. First, the dataset (401) used to train M is browsed through the identifier (503-1), and data items (402, 403) for labels included in the standard objective label list are collected to construct D. If there are any standard objective labels for which data could not be obtained in this way, the dataset storage (400) is searched for data items that are identical to the standard objective label in the correct answer data (403-2), and these are added to D. After constructing D, a table L is also constructed that associates each label in the standard objective label list with a class index.

[0157] This step involves considering various factors necessary for training M, such as equalizing the number of data points per class, determining the appropriate number of data points per class for the scale of M, and splitting D into training, validation, and test data.

[0158] If an "unknown" node is generated in step (S2002), a label not included in the standard target label is randomly selected, and the data is collected and assigned to the "unknown" class to constitute dataset D. By training the intelligent model to include data not included in the class corresponding to the standard target label in the "unknown" class, the probability of false positives can be reduced, improving the classification accuracy of the intelligent model.

[0159] Next, M2 is generated by training M1 using D (S2004). Various training methods can be applied depending on the type of intelligent model, and as mentioned earlier, such code is secured through the intelligent model utilization code storage (700) of the intelligent repository (203). This step can be performed by inputting D and M1 into the previously secured "training" code and driving it.

[0160] At this time, M2 and D are registered in the intelligent repository (203). The intelligent model data (502) and intelligent model metadata (503) of M2 must be properly described. The intelligent model identifier (502-1) is newly generated and registered, and the model parameter values ​​(502-3), correct answer label list (503-2), training history (503-4), and performance evaluation information (503-5) must be recorded with appropriate information. The dataset identifier (503-1) is the dataset identifier (402-1) of D. The base model (503-3) is the intelligent model identifier (502-1) of M. D is recorded by registering a new dataset identifier (401-1) and storing the correct answer data list (401-2) that constitutes D.

[0161] By performing steps (S2002) to (S2004), it is possible to reduce the size of the intelligent model and improve its accuracy.

[0162] Next, a dataset D1 is constructed (S2005) to be used to optimize M2 to the intelligence requirements, based on the target data included in the intelligence requirements profile (110). D1 consists of pairs of data items included in the raw data (110-3) and data annotations (110-4) corresponding to each item. The correct answers for the data annotations must be converted to a standard vocabulary through the label dictionary (300), and then the class index must be obtained through L constructed in step (S2003), after which the correct answers for each raw data item must be obtained.

[0163] Next, M2 is trained with D1 to generate M3 (S2006). This can be done by inputting D1 and M2 into the "training" code reserved in the intelligent model utilization code storage (700), as in step (S2004), and driving it.

[0164] Register M3 and D1 in the intelligent repository (203). The intelligent model data (502) and intelligent model metadata (503) for M3 must be properly documented. The intelligent model identifier (502-1) must be newly generated and registered, and the model parameter values ​​(502-3), correct answer label list (503-2), training history (503-4), and performance evaluation information (503-5) must be recorded with appropriate information. The dataset identifier (503-1) should contain the dataset identifier (402-1) for D1. The base model (503-3) should contain the intelligent model identifier (502-1) for M2. D1 is recorded by registering a new dataset identifier (401-1) and storing the correct answer data list (401-2) that constitutes D1.

[0165] In the initial selection in step (S2001)1, it may not be possible to find an intelligent model in the intelligent model storage that satisfies the task details (110-1) described in the intelligence requirement profile. In this case, the intelligent manager (201) can select and utilize an intelligent model type (601) stored in the intelligent model type dictionary (600) whose task identifier (601-3) is the same as the task identifier in the task details (110-1). After selecting an intelligent model type (601), the model type structure details (601-2) are restored to generate an initial model BM with empty model parameter values. Subsequently, BM can be used in place of M. Since BM is an empty model that has not been trained, it will only in step (S2004) generate a model that performs its intended function. The subsequent process is as described above.

[0166] The following details how to modify the intelligent requirements profile for data security.

[0167] In step (S1002), if the edge server (150) fails to generate an intelligent model, it transmits the intelligent request profile to the cloud server and entrusts the intelligent generation task to it. At this time, the intelligent administrator (201) of the edge server (150) performs data protection by modifying the intelligent request profile considering the scope of data disclosure and transmitting it to the cloud server.

[0168] Figure 4 shows an example of an intelligence requirements profile, where the scope of publication is described as "regional" and "global." In this case, data from customers, edge server owners, or service providers can be protected by applying rules to prevent edge servers from transmitting data limited to a specific region to the cloud server. The intelligence requirements profile that edge servers transmit to the cloud server includes: 1) all target data with a "global" scope, 2) correct labels for target data with a "regional" scope, and 3) a list of target correct labels. The edge server stores and saves the target data with a "regional" scope for future regional optimization of the intelligent model.

[0169] In further embodiments, the scope of disclosure can be specified in multiple stages. As shown in Figure 3, intermediate servers can be placed in multiple stages between the terminal and the final cloud server, so in such cases, the scope of disclosure can be precisely defined and described to determine which step's server the data can be transmitted to. In yet another embodiment, the scope of disclosure can be specified automatically. For example, a detector that can determine whether a person is present in a photograph or video can be provided, and detection can be performed on raw data included in the intelligent request profile, and the scope of disclosure for any data containing a person can be set to "region". In this way, the entity managing or using the system according to the present invention can specify a particular object to set the scope of disclosure and have the edge server automatically process the data security function.

[0170] Figure 13 shows an example of how the edge server has modified the intelligent requirements profile from Figure 4 based on the data exposure scope.

[0171] The data related to img02.jpg, whose public access scope is "Region," was removed from the intelligence requirement profile, and "Cup," the correct answer label for img02.jpg, was added as the target correct answer label. This was because there was no raw data corresponding to "Cup" in the profile. In this case, the public access scope item did not need to be transmitted to the cloud server, so it was removed from the intelligence requirement profile.

[0172] The following section provides a detailed explanation of how to optimize intelligent models on edge servers.

[0173] In step (S1002), the edge server self-stored the raw data (110-3) and its data annotations (110-4) whose publication scope is "regional" from the intelligence request profile (110). Based on the self-stored data items in this way, it constructed dataset D2. D2 consists of pairs of raw data items and the correct answers in the data comments.

[0174] The edge server (150) receives the intelligent model M3 and generates the final intelligent model M4 requested by the initial intelligence requirement profile (110) by training it using D2. The training method is the same as in steps (S2004) and (S2006). This allows data that could not be used for intelligent model optimization on the cloud server due to its limited public access to be applied to the performance optimization of the intelligent model.

[0175] D2 and M4 will also be registered in the intelligent repository (203) using the same method as D, D1, M2, and M3 were registered earlier.

[0176] The following provides a detailed explanation of the quality control methods for intelligent models.

[0177] The quality history (503-6) information included in the intelligent model metadata (503) can be used as reference material for selecting high-quality intelligent models that have no history of problems. If the quality of an intelligent model is significantly low or if it poses a critical risk, its use in critical tasks can be prevented.

[0178] For example, if model M generates an error and causes a problem under specific circumstances, information describing that situation is transmitted to the edge server or cloud server via the intelligent repository interface (204). During transmission, the edge server and cloud server add this information to the quality history (503-6) item in the intelligent model metadata (503) of model M. In the future, by viewing this item, the quality of the intelligent model can be predicted. If an intelligent model M experiences a serious performance degradation or problem under specific circumstances, it is possible to identify intelligent models that are potentially prone to problems by finding models identical or similar to M.

[0179] Models identical to M can be found by comparing the intelligent model identifiers (502-1) included in the intelligent model data (502).

[0180] In one embodiment of the present invention, a model similar to M can be found as follows.

[0181] 1) The base model (503-3) described in the intelligent model metadata (503) of M is the intelligent model used to generate M, and is therefore judged to be a similar model. The base model of M may be generated from another base model. In this way, similar models of M can be found by sequentially referencing the base models of intelligent models.

[0182] 2) Similar models can be found by measuring the similarity of the intelligence model data between two intelligence models. By comparing the model type (502-2), training dataset (503-1), correct answer label list (503-2), base model (503-3), training history (503-4), etc., of the two intelligence models, they can be judged as similar models if they are similar.

[0183] While such similarities between data do not necessarily prove similarities in the operating characteristics of the two models, they can serve as clues to predict the likelihood of problems occurring.

[0184] The structure and contents of the intelligent repository will be described in detail below with reference to Tables 4 to 8.

[0185] Table 4 shows one example of the intelligence requirements profile (210).

[0186] Table 5 shows one example of a label dictionary (300).

[0187] Table 6 shows one example of the intelligent model typology dictionary (600).

[0188] Table 7 shows one example of the intelligent model storage (500).

[0189] Table 8 shows one example of the intelligent model-utilizing code storage (700). [Table 4] [Table 5] [Table 6] [Table 7] [Table 8]

[0190] The intelligence requirements profile (210) in [Table 4] indicates that it requests an intelligent model that can receive an image input and detect seven object classes.

[0191] To select an intelligent model that satisfies these requirements, the task details (110-4) of the intelligence requirements profile are compared with the task details (502-4) of each intelligent model in the intelligent model storage (500), and identical ones are selected. Referring to [Table 7], it can be seen that IM000011 satisfies these conditions.

[0192] Looking at the intelligent model type identifier (502-2) of intelligent model IM000009, we see that the model structure is IMT00003. Looking at the intelligent model type dictionary (600), we can see that the structural details of this model are formally described in alexnet01.onnx, and that it can be used for classification tasks. The alexnet01.ONNX details, when used with a deep learning framework compatible with the ONNX structure, can be used to generate and utilize the initial intelligent model created with that model structure before it is trained through reconstruction.

[0193] The training history (503-4) for IM000009 allows you to view the settings for various parameters used during training, such as epoch, batch size, and learning rate.

[0194] Looking at the performance evaluation information (503-5) for IM000009, we can see that it achieved a recall performance of 0.992 and a precision performance of 0.87 on the DS000001 dataset. For IM000015, using the same dataset, the recall performance was 0.96 and the precision performance was 0.89. Performance can be compared by comparing the same performance values ​​with intelligent models that have the same task details.

[0195] The quality history (503-6) for IM000009 shows a report from 2021-07-03 with a severe status and a URL for related information. This indicates that the intelligent model in question has previously caused serious problems.

[0196] The intelligent model utilization code storage (700) contains code CD000001, which can perform inference on the IM000009 model, and code CD000002, which can perform training. These are concrete examples of containers, which store identifiers (e.g., imcloud / imt00003:inference) and driving scripts (e.g., script001.bash). These codes should be used when generating, optimizing, and utilizing intelligent models using the IM000009 model.

[0197] Figure 14 is a block diagram showing the structure of an edge server in one embodiment of the present invention.

[0198] Referring to Figure 14, an edge server according to one embodiment of the present invention includes a communication unit (21) that communicates with user terminals and other servers, a storage unit (22) that stores data for intelligent model generation, a model generation unit (23) that generates an intelligent model in response to an intelligent model generation request, and an adjustment unit (24) that adjusts the generated intelligent model.

[0199] In this case, if the model generation unit fails to generate the intelligent model, the communication unit (24) may request the cloud server to generate the intelligent model and receive the intelligent model generated by the cloud server.

[0200] In this case, the cloud server may include a first cloud server and a second cloud server having a larger capacity than the first cloud server.

[0201] In this case, if the first cloud server fails to generate the intelligent model, it may request the second cloud server to generate the intelligent model.

[0202] In this case, the intelligent model generation request may include a task identifier, raw data, annotations, data disclosure scope, and target labels.

[0203] In this case, the model generation unit (23) may select a basic intelligence model based on the intelligence model generation request, modify the label list of the basic intelligence model to correspond to the target label list, and perform training on the modified intelligence model.

[0204] In this case, the communication unit (21) may transmit the raw data to the cloud server based on the data disclosure scope.

[0205] In this case, the adjustment unit (24) may adjust the intelligent model using raw data that has not been transmitted to the cloud server.

[0206] Figure 15 is a block diagram showing the structure of a cloud server according to one embodiment of the present invention.

[0207] Referring to Figure 15, a cloud server according to one embodiment of the present invention includes a communication unit (31) that receives intelligent model generation requests from edge servers, a storage unit (32) that stores data for intelligent model generation, and a model generation unit (33) that generates an intelligent model corresponding to the intelligent model generation request, the intelligent model generation request may include a task identifier, raw data, annotations, data disclosure scope, and target labels.

[0208] In this case, if the model generation unit fails to generate the intelligent model, the communication unit (31) may request other cloud servers to generate the intelligent model.

[0209] In this case, the raw data of the intelligent model generation request may be transmitted by the edge server based on the data disclosure scope.

[0210] Figure 16 shows the configuration of a computer system according to an embodiment.

[0211] The edge server and cloud server according to the embodiment can be realized as a computer system (1000) such as a computer-readable recording medium.

[0212] The computer system (1000) may include one or more processors (1010), memory (1030), user interface input devices (1040), user interface output devices (1050), and storage (1060) that communicate with each other via a bus (1020). Furthermore, the computer system (1000) may further include a network interface (1070) connected to a network (1080). The processor (1010) may be a central processing unit or a semiconductor device that executes programs or processing instructions stored in the memory (1030) or storage (1060). The memory (1030) and storage (1060) may be storage media that include at least one of the following: volatile media, non-volatile media, separable media, non-separable media, communication media, or information transmission media. For example, the memory (1030) may include ROM (1031) or RAM (1032).

[0213] The specific executions described herein are examples and do not in any way limit the scope of the invention. For the sake of brevity of the specification, descriptions of conventional electronic configurations, control systems, software, and other functional aspects of said systems may be omitted. Furthermore, the connections of lines or connecting members between components shown in the drawings are illustrative examples of functional and / or physical or circuit connections and may be substituted or shown as various additional functional, physical, or circuit connections in actual devices. Moreover, components that are not necessarily required to apply the invention may not be necessary unless specifically mentioned, such as "essential" or "important."

[0214] Therefore, the concept of the present invention should not be limited to the embodiments described above, and not only the claims described later, but also all scopes equivalent to or equivalently modified thereunder, can be said to fall within the scope of the concept of the present invention. [Explanation of Symbols]

[0215] 1000: Computer System 1010: Processor 1020: Bus 1030: Memory 1031: Rom 1032: Ram 1040: User Interface Input Device 1050: User Interface Output Device 1060: Storage 1070: Network Interface 1080: Network

Claims

1. A communication unit that communicates with the terminal or a higher-level server, A storage unit containing data for generating intelligent models, A model generation unit that generates an intelligent model based on an intelligence requirements profile, An adjustment unit for adjusting the aforementioned intelligent model, Includes, The intelligence requirement profile includes the task details of an intelligent model that responds to an intelligent model generation request and the target data used to train the intelligent model. The aforementioned target data includes raw data and the data disclosure scope corresponding to the said raw data, The model generation unit uses the data disclosure scope to divide the raw data into public raw data provided to at least one upstream server or non-public raw data not provided to the upstream server. An intermediate server characterized in that the model generation unit removes the non-public raw data and data annotations corresponding to the non-public raw data from the intelligence requirement profile, adds target correct answer labels corresponding to the data annotations to the intelligence requirement profile to generate a modified intelligence requirement profile, and provides the intelligence requirement profile to the higher-level server.

2. The aforementioned model generation unit, The intelligent model generated based on the modified intelligence request profile is received from the aforementioned upper-level server. The adjustment unit is The intermediate server according to claim 1, characterized in that it adjusts the intelligent model generated from the higher-level server based on the aforementioned non-public raw data.

3. The aforementioned model generation unit, The intermediate server according to claim 1, characterized in that it selects a basic intelligence model from among the intelligence models stored in the storage unit based on the intelligence requirement profile, and generates the intelligence model based on the basic intelligence model.

4. The aforementioned basic intelligence model is, The intermediate server according to claim 3, characterized in that it is selected based on the similarity between the label list of the basic intelligence model and the target label list included in the intelligence requirements profile.

5. The aforementioned model generation unit, The intermediate server according to claim 3, characterized in that it transforms the label list of the basic intelligence model to correspond to the target label list included in the intelligence requirement profile, performs training on the transformed intelligence model, and generates the intelligence model.

6. Learning for the modified intelligence model is A first learning step using the dataset stored in the storage unit, The intermediate server according to claim 5, characterized in that it is performed using a second learning step that uses raw data included in the intelligence requirements profile.

7. The aforementioned already stored dataset is The intermediate server according to claim 6, characterized in that it is composed of data corresponding to the target label list from among the data stored in the storage unit.

8. In a method for generating intelligent models performed on a server, The steps include receiving a request to generate an intelligent model for the terminal, The steps include generating an intelligent model based on an intelligence requirements profile, The steps include adjusting the aforementioned intelligent model, The intelligence requirements profile includes the task details of the intelligence model and the target data used to train the intelligence model. The aforementioned target data includes raw data and the data disclosure scope corresponding to the said raw data, In the step of generating an intelligence model based on the intelligence requirements profile, Using the aforementioned data disclosure scope, the raw data is divided into public raw data provided to at least one upstream server or non-public raw data not provided to the upstream server. A method for generating an intelligent model, characterized by removing the non-public raw data and data annotations corresponding to the non-public raw data from the intelligence requirement profile, adding target correct answer labels corresponding to the data annotations to the intelligence requirement profile to generate a modified intelligence requirement profile, and providing the intelligence requirement profile to the higher-level server.

9. The step of generating the aforementioned intelligent model is: The method for generating an intelligent model according to claim 8, characterized in that, based on the intelligence requirement profile, a basic intelligent model is selected from the intelligent models stored in the storage unit, and the intelligent model is generated based on the basic intelligent model.

10. The aforementioned basic intelligence model is, The method for generating an intelligent model according to claim 9, characterized in that it is selected based on the similarity between the label list of the basic intelligent model and the target label list included in the intelligent requirements profile.

11. The storage unit stores intelligent model metadata corresponding to each of the stored intelligent models, The intelligent model generation method according to claim 9, characterized in that the intelligent model metadata includes the training history, performance evaluation information, or quality history of the intelligent model.

12. The training history includes change history data for the training parameters of the intelligent model, The performance evaluation information includes the evaluation dataset and performance data of the intelligent model. The intelligent model generation method according to claim 11, characterized in that the quality history includes problem data that occurred during the intelligent model utilization process.

13. The step of generating the aforementioned intelligent model is: The method for generating an intelligent model according to claim 11, characterized in that a basic intelligent model is selected from the intelligent models stored in the storage unit based on the intelligent model metadata.

Images