Artificial intelligence system using intensive learning and control method thereof

Abstract

The present invention discloses an artificial intelligence system using focused learning and a method for controlling the same. That is, the present invention selects a plurality of operable terminals from among all terminals based on terminal-specific model learning status information, terminal-specific CPU usable capacity, terminal-specific GPU usable capacity, and raw data characteristic information for all raw data, divides the entire raw data into a plurality of sub-raw data, allocates the plurality of sub-raw data to each of the selected operable terminals, performs AI-based learning based on the sub-raw data allocated to each of the operable terminals to generate sub-results, collects the plurality of sub-results based on the learning results from the plurality of operable terminals in the first stage at the management terminal, performs AI-based learning independently based on the collected plurality of sub-results and the plurality of raw data to generate additional sub-results, merges the collected plurality of sub-results and the generated additional sub-results to generate and transmit a final result, and performs verification and test functions based on the final result transmitted from the management terminal at the integration terminal to generate learning data including verification eigenvalues, absolute values, minimum values, maximum values, etc. for the entire raw data, thereby reducing the time required for learning through intensive learning of the AI-based model and improving the operational efficiency of the entire system.

Description

Technology Field

[0001] The present invention relates to an artificial intelligence system using intensive learning and a control method thereof, and more specifically, to an artificial intelligence system using intensive learning and a control method thereof, wherein, for all raw data, a plurality of operable terminals are selected from among all terminals based on terminal-specific model learning status information, terminal-specific CPU usable capacity, terminal-specific GPU usable capacity, and raw data characteristic information, and the entire raw data is divided into a plurality of sub-raw data, then the plurality of sub-raw data are each assigned to the selected plurality of operable terminals, and artificial intelligence-based learning is performed based on the sub-raw data assigned to each of the plurality of operable terminals to generate sub-results, a management terminal first collects a plurality of sub-results based on the learning results from the plurality of operable terminals, and performs artificial intelligence-based learning independently based on the first collected plurality of sub-results and the plurality of raw data to generate additional sub-results, and merges the first collected plurality of sub-results and the generated additional sub-results to generate and transmit a final result, and an integrated terminal performs verification functions and test functions based on the final result transmitted from the management terminal to generate learning data including verification eigenvalues, absolute values, minimum values, maximum values, etc. for the entire raw data. Background Technology

[0002] Deep learning is a technology used to classify or cluster objects or data, and is a type of machine learning that trains computers to classify objects just as the human brain distinguishes objects.

[0003] Training models used in such deep learning takes a lot of time. Prior art literature

[0004] Korean Registered Patent No. 10-2566928 [Title: Electronic device for generating a training set for performing reinforcement learning of a deep learning model for determining user intent and method of operation thereof] The problem to be solved

[0005] The objective of the present invention is to provide an artificial intelligence system using intensive learning and a control method thereof, which, for all raw data, selects a plurality of operable terminals from among all terminals based on terminal-specific model learning status information, terminal-specific CPU usable capacity, terminal-specific GPU usable capacity, and raw data characteristic information, divides the entire raw data into a plurality of sub-raw data, allocates the plurality of sub-raw data to each of the selected operable terminals, performs artificial intelligence-based learning based on the sub-raw data allocated to each of the operable terminals to generate sub-results, collects the plurality of sub-results based on the learning results from the plurality of operable terminals in a first step at a management terminal, performs artificial intelligence-based learning independently based on the collected plurality of sub-results and the plurality of raw data to generate additional sub-results, merges the collected plurality of sub-results and the generated additional sub-results to generate and transmit a final result, and performs verification and test functions based on the final result transmitted from the management terminal at an integrated terminal to generate learning data including verification eigenvalues, absolute values, minimum values, maximum values, etc. for the entire raw data. means of solving the problem

[0006] An artificial intelligence system using intensive learning according to an embodiment of the present invention classifies all raw data into raw data for high-performance computing and raw data for low-performance computing based on characteristic information for each raw data, selects a plurality of operable terminals among all high-performance terminals and low-performance terminals managed by an integrated terminal, divides one or more raw data classified as high-performance computing raw data and at least one raw data classified as low-performance computing raw data into a plurality of sub-raw data based on the CPU usable capacity or GPU usable capacity per terminal of the selected plurality of operable terminals, assigns the divided plurality of sub-raw data to the plurality of high-performance terminals and the plurality of low-performance terminals respectively, and provides the plurality of sub-raw data assigned to the plurality of high-performance terminals and the plurality of low-performance terminals respectively, information on a pre-selected supervised learning model, and information on a pre-set epoch to the plurality of high-performance terminals and the plurality of low-performance terminals respectively; It may include a plurality of high-performance terminals and a plurality of low-performance terminals that perform artificial intelligence-based learning based on a plurality of sub-raw data provided from the integrated terminal, information about any one of the supervised learning models, and information about the epoch, respectively generate a sub-result as a learning result, and transmit the sub-result to a management terminal whenever a sub-result is generated by the learning performance according to the execution of the epoch.

[0007] As an example related to the present invention, the management terminal interacts with the plurality of high-performance terminals or the plurality of low-performance terminals to collect sub-results generated according to the performance of learning at the plurality of high-performance terminals or the plurality of low-performance terminals, and according to the entire epoch, after the collection of multiple sub-results related to the plurality of raw data from the plurality of high-performance terminals and the plurality of low-performance terminals is completed, performs additional learning based on artificial intelligence based on the plurality of raw data, the multiple sub-results for each of the collected raw data, any one of the supervised learning models, and information regarding the epoch, generates additional sub-results related to the corresponding plurality of raw data based on the additional learning results, merges the collected sub-results and the generated additional sub-results to generate a final result related to the corresponding specific object that is the subject of learning, and transmits the generated final result to the integration terminal, and the integration terminal performs a verification function and a test function respectively on the final result transmitted from the management terminal, and can generate and store learning data for the entire raw data related to the specific object based on the final result, the result of the performed verification function, and the result of the performed test function. there is.

[0008] A control method for an artificial intelligence system using intensive learning according to an embodiment of the present invention comprises: a step of classifying all raw data into raw data for high-performance computing and raw data for low-performance computing based on characteristic information for each raw data by an integrated terminal; a step of selecting a plurality of operable terminals from among all high-performance terminals and low-performance terminals managed by the integrated terminal by the integrated terminal; a step of dividing one or more raw data classified as high-performance computing raw data and at least one raw data classified as low-performance computing raw data into a plurality of sub-raw data by the integrated terminal based on the CPU usable capacity or GPU usable capacity per terminal of the plurality of high-performance terminals and the plurality of low-performance terminals selected as operable terminals by the integrated terminal; and a step of allocating the divided plurality of sub-raw data to each of the plurality of high-performance terminals and the plurality of low-performance terminals selected as operable terminals by the integrated terminal. A step of providing, by the integrated terminal, to each of the plurality of high-performance terminals and the plurality of low-performance terminals, a plurality of sub-raw data assigned to each of the plurality of high-performance terminals and the plurality of low-performance terminals, information regarding a predetermined supervised learning model, and information regarding a predetermined epoch; a step of, by each of the plurality of high-performance terminals and the plurality of low-performance terminals, performing AI-based learning based on the plurality of sub-raw data provided from the integrated terminal, information regarding the predetermined supervised learning model, and information regarding the epoch, thereby generating sub-results as learning results, and transmitting the corresponding sub-results to a management terminal whenever a sub-result is generated by learning performance according to the execution of an epoch; a step of, by the management terminal, collecting sub-results generated according to learning performance at the plurality of high-performance terminals or the plurality of low-performance terminals in conjunction with the plurality of high-performance terminals or the plurality of low-performance terminals.The method may include the steps of: by the management terminal, after the collection of multiple sub-results related to the multiple raw data from the multiple high-performance terminals and the multiple low-performance terminals according to the entire epoch, is completed, performing additional AI-based learning based on the multiple raw data, the multiple sub-results for each collected raw data, any one supervised learning model, and information regarding the epoch, and generating additional sub-results related to the multiple raw data based on the additional learning results; by the management terminal, merging the collected sub-results and the generated additional sub-results to generate a final result related to the specific object being learned, and transmitting the generated final result to the integration terminal; by the integration terminal, performing a verification function and a test function, respectively, on the final result transmitted from the management terminal; and by the integration terminal, generating and storing training data for the entire raw data related to the specific object based on the final result, the result of the performed verification function, and the result of the performed test function.

[0009] As an example related to the present invention, the step of classifying the entire raw data into raw data for high-performance computation and raw data for low-performance computation may include: a process of performing a preset Fourier transform function for each of the class-labeled entire images when the data type of the raw data is an image; a process of performing filtering through a preset high-pass filter for each of the Fourier transformed entire images; a process of calculating an image-specific frequency average value for each of the filtered entire images; a process of calculating an entire image frequency average value based on the calculated image-specific frequency average value; a process of identifying one or more image frequency average values ​​among the calculated image frequency average values ​​that are greater than or equal to a preset threshold value than the calculated entire image frequency average value; a process of classifying one or more images among the entire images that correspond to one or more image frequency average values ​​among the entire images that are greater than or equal to a preset threshold value than the identified calculated entire image frequency average value as images for high-performance computation; and a process of classifying the remaining images among the entire images, excluding the one or more images classified as images for high-performance computation, as images for low-performance computation.

[0010] As an example related to the present invention, the step of classifying the entire raw data into raw data for high-performance computation and raw data for low-performance computation may include: a process of performing a preset Fast Fourier Transform function for each of the entire class-labeled sounds when the data type of the raw data is sound; a process of calculating a sound-specific resolution for each of the entire sounds that have undergone Fast Fourier Transform; a process of calculating an average value of the entire sound resolution based on the calculated sound-specific resolution; a process of identifying one or more sound resolutions among the calculated sound-specific resolutions that are greater than or equal to a preset threshold value than the calculated average value of the entire sound resolution; a process of classifying one or more sounds among the entire sounds that correspond to one or more sound resolutions among the entire sounds that are greater than or equal to a preset threshold value than the identified calculated average value of the entire sound resolution as sounds for high-performance computation; and a process of classifying the remaining sounds among the entire sounds, excluding the one or more sounds classified as sounds for high-performance computation, as sounds for low-performance computation.

[0011] As an example related to the present invention, the step of classifying the entire raw data into raw data for high-performance computation and raw data for low-performance computation comprises: a process of calculating a jump size per text or a time complexity per text by performing a preset Boyer-Moore algorithm on each of the class-labeled entire text data when the data type of the raw data is text; a process of calculating an average jump size of the entire text data or an average time complexity of the entire text data based on the calculated jump size per text data or time complexity per text data; and a process of identifying one or more text data jump sizes or one or more text data time complexes among the calculated jump sizes per text data or the calculated time complexes per text data that are greater than or equal to another preset threshold value than the calculated average jump size per text data or the calculated average time complexity per text data. Among the above-mentioned total text data, the process may include: classifying one or more text data corresponding to one or more text data jump sizes or one or more text data time complexes that are greater than or equal to another threshold value pre-set than the identified calculated total text data jump size average value or the calculated total text data time complexity average value as high-performance computation text data; and among the above-mentioned total text data, the process may include classifying the remaining raw data, excluding one or more raw data classified as high-performance computation text data, as low-performance computation text data.

[0012] As an example related to the present invention, the step of selecting a plurality of operable terminals among all high-performance terminals and low-performance terminals managed by the integrated terminal may include: a process of selecting a plurality of high-performance terminals as operable terminals among all high-performance terminals managed by the integrated terminal that are in a state where a learning function is not being performed using a learning model and whose available GPU capacity is greater than or equal to a preset threshold value; and a process of selecting a plurality of low-performance terminals as operable terminals among all low-performance terminals managed by the integrated terminal that are in a state where a learning function is not being performed using a learning model and whose available CPU capacity is greater than or equal to another preset threshold value. Effects of the invention

[0013] The present invention selects a plurality of operable terminals from among all terminals based on terminal-specific model learning status information, terminal-specific CPU usable capacity, terminal-specific GPU usable capacity, and raw data characteristic information for all raw data, divides the entire raw data into a plurality of sub-raw data, allocates the plurality of sub-raw data to each of the selected operable terminals, performs AI-based learning based on the sub-raw data allocated to each of the operable terminals to generate sub-results, collects the plurality of sub-results based on the learning results from the plurality of operable terminals in the first stage at a management terminal, performs AI-based learning independently based on the collected plurality of sub-results and the plurality of raw data to generate additional sub-results, merges the collected plurality of sub-results and the generated additional sub-results to generate and transmit a final result, and performs verification and test functions based on the final result transmitted from the management terminal at an integrated terminal to generate learning data including verification eigenvalues, absolute values, minimum values, maximum values, etc. for the entire raw data, thereby reducing the learning time required through intensive learning of an AI-based model and improving the operational efficiency of the entire system. Brief explanation of the drawing

[0014] FIG. 1 is a block diagram showing the configuration of an artificial intelligence system using focused learning according to an embodiment of the present invention. FIGS. 2 and 3 are flowcharts illustrating a control method of an artificial intelligence system using focused learning according to an embodiment of the present invention. Specific details for implementing the invention

[0015] It should be noted that the technical terms used in this invention are used merely to describe specific embodiments and are not intended to limit the invention. Furthermore, unless specifically defined otherwise in this invention, the technical terms used in this invention should be interpreted in the sense generally understood by those skilled in the art to which this invention pertains, and should not be interpreted in an overly broad or overly narrow sense. Additionally, if a technical term used in this invention is an incorrect technical term that fails to accurately express the concept of the invention, it should be replaced with a technical term that can be correctly understood by those skilled in the art. Moreover, general terms used in this invention should be interpreted according to their prior definitions or the context, and should not be interpreted in an overly narrow sense.

[0016] Furthermore, singular expressions used in the present invention include plural expressions unless the context clearly indicates otherwise. Terms such as "composed of" or "comprising" in the present invention should not be interpreted as necessarily including all of the various components or steps described in the invention, and should be interpreted as meaning that some of the components or steps may not be included, or that additional components or steps may be included.

[0017] Additionally, terms including ordinal numbers, such as first, second, etc., used in the present invention may be used to describe components, but the components should not be limited by the terms. The terms are used solely for the purpose of distinguishing one component from another. For example, without departing from the scope of the present invention, the first component may be named the second component, and similarly, the second component may be named the first component.

[0018] Hereinafter, preferred embodiments according to the present invention will be described in detail with reference to the attached drawings. Identical or similar components are given the same reference number regardless of the drawing symbols, and redundant descriptions thereof will be omitted.

[0019] Furthermore, in describing the present invention, detailed descriptions of related prior art are omitted if it is determined that such descriptions could obscure the essence of the invention. Additionally, it should be noted that the attached drawings are intended only to facilitate an understanding of the concept of the present invention and should not be interpreted as limiting the concept of the present invention.

[0020] FIG. 1 is a block diagram showing the configuration of an artificial intelligence system (10) using intensive learning according to an embodiment of the present invention.

[0021] As illustrated in FIG. 1, the artificial intelligence system (10) using intensive learning is composed of a plurality of high-performance terminals (100), a plurality of low-performance terminals (200), a management terminal (300), and an integrated terminal (400). Not all components of the artificial intelligence system (10) using intensive learning illustrated in FIG. 1 are essential components, and the artificial intelligence system (10) using intensive learning may be implemented with more components than those illustrated in FIG. 1, or with fewer components.

[0022] The artificial intelligence system (10) using the above intensive learning includes a cloud system, a distributed system, a parallel system, etc.

[0023] The high-performance terminal (100), the low-performance terminal (200), the management terminal (300), and the integrated terminal (400) are a smartphone, portable terminal, mobile terminal, foldable terminal, personal digital assistant (PDA), portable multimedia player (PMP) terminal, telematics terminal, navigation terminal, personal computer, laptop computer, slate PC, tablet PC, ultrabook, wearable device (including, for example, smartwatch, smart glass, head-mounted display, etc.), Wibro terminal, IPTV terminal, smart TV, digital broadcasting terminal, AVN terminal, A / V system, flexible It can be applied to various terminals, such as flexible terminals and digital signage devices.

[0024] Additionally, the high-performance terminal (100), the low-performance terminal (200), the management terminal (300), and the integrated terminal (400) may be configured as servers.

[0025] In addition, when the high-performance terminal (100), the low-performance terminal (200), the management terminal (300), and the integrated terminal (400) are configured as a server, the server may be implemented in the form of a web server, a database server, a proxy server, etc. Furthermore, one or more of a network load balancing mechanism or various software that enables the server to operate on the Internet or another network may be installed on the server, thereby enabling it to be implemented as a computerized system. In addition, the network may be an HTTP network, a private line, an intranet, or any other network. Furthermore, the connection between the high-performance terminal (100), the low-performance terminal (200), the management terminal (300), and the integrated terminal (400) may be connected via a secure network to prevent data from being attacked by any hacker or other third party. In addition, the server may include a plurality of database servers, and such database servers may be implemented in a manner where they are connected separately from the server through any type of network connection, including a distributed database server architecture.

[0026] Each of the above high-performance terminal (100), the above low-performance terminal (200), the above management terminal (300), and the above integrated terminal (400) may include a communication unit (not shown) for performing communication functions with other terminals, a storage unit (not shown) for storing various information and programs (or applications), a display unit (not shown) for displaying various information and program execution results, a voice output unit (not shown) for outputting voice information corresponding to the various information and program execution results, and a control unit (not shown) for controlling various components and functions of each terminal.

[0027] The above plurality of high-performance terminals (100) communicate with other high-performance terminals (100), the above plurality of low-performance terminals (200), the above management terminal (300), the above integration terminal (400), etc. At this time, the high-performance terminal (100) may be a terminal that satisfies the CPU type and capacity, GPU type and capacity, RAM type and capacity, etc., which are pre-set based on the CPU type and capacity, GPU type and capacity, RAM type and capacity, etc. of the high-performance terminal (100).

[0028] In addition, the high-performance terminal (100) is registered (or subscribed) as a terminal for performing intensive learning functions on the integrated terminal (400) through linkage with the integrated terminal (400), and registers a terminal specification including a pre-set CPU type and capacity, GPU type and capacity, RAM type and capacity, etc. based on the CPU type and capacity, GPU type and capacity, RAM type and capacity, etc. of the high-performance terminal (100).

[0029] In addition, the high-performance terminal (100) registers a consent form for the use of personal information, a consent form for the collection of personal information, a consent form for the provision of personal information, a consent form for the processing of unique identification information, etc., in the form of an electronic document pre-set in the high-performance terminal (100) in order to use the functions (or services) provided by the integrated terminal (400) through linkage with the integrated terminal (400).

[0030] In addition, the high-performance terminal (100) accumulates points provided by the integrated terminal (400) by performing an intensive learning function related to a specific object using raw data provided by the integrated terminal (400). The points accumulated in this way can be used like cash through a shopping mall server (not shown), etc.

[0031] In addition, the high-performance terminal (100) can display (or receive) statistical information regarding the learning results using artificial intelligence performed at the request of the integrated terminal (400) through a dedicated app and / or dedicated website provided by the integrated terminal (400).

[0032] Additionally, the high-performance terminal (100) receives a plurality of sub-raw data allocated to the high-performance terminal (100) provided (or transmitted) from the integrated terminal (400), information about any one of the supervised learning models (or at least one supervised learning model), information about the epoch, etc. At this time, the plurality of sub-raw data may be allocated (or assigned) according to the available GPU capacity of the high-performance terminal (100) among a plurality of raw data related to a specific object to be learned.

[0033] Additionally, the high-performance terminal (100), according to information regarding any one of the received supervised learning models (or at least one supervised learning model), performs learning (or artificial intelligence / machine learning / deep learning) by using the received (or each assigned to each high-performance terminal) multiple sub-raw data as input values ​​for the corresponding artificial intelligence-based model (or learning model), and generates sub-results related to the corresponding sub-raw data based on the learning results (or artificial intelligence results / machine learning results / deep learning results). Here, the sub-result (or sub-learning result / first sub-result) includes information regarding the corresponding sub-raw data (or including an index, ID, etc. regarding the corresponding raw data), the number of epochs (or the number of epochs), the results of the learning execution, etc. At this time, the high-performance terminal (100) performs a learning function for the corresponding multiple sub-raw data using a GPU (Graphics Processing Unit) among the various components constituting the high-performance terminal (100).

[0034] In addition, the high-performance terminal (100) transmits the sub-results generated each time an epoch of learning is achieved, identification information of the high-performance terminal (100), etc., to the management terminal (300). Here, the identification information of the high-performance terminal (100) includes an Agent Server ID, MDN (Mobile Directory Number), mobile IP, mobile MAC, unique SIM (subscriber identity module) card information, serial number, etc. At this time, the Agent Server ID may be an ID (or unique ID) associated with a plurality of daemon programs executed on each terminal (100, 200, 300, 400) (e.g., computers, etc.).

[0035] In the embodiment of the present invention, it is mainly described that a sub-result related to the sub-raw data is generated by the learning result through the specific supervised learning model selected earlier, but it is not limited thereto. Each of the plurality of high-performance terminals (100) may perform learning in relation to the sub-raw data for each of the plurality of supervised learning models that are pre-set so as to compare the learning results through the plurality of supervised learning models, generate a sub-result for each supervised learning model according to the learning result, and transmit the generated sub-result for each supervised learning model to the management terminal (300).

[0036] The low-performance terminal (200) communicates with the plurality of high-performance terminals (100), other low-performance terminals (200), the management terminal (300), the integrated terminal (400), etc. At this time, the low-performance terminal (200) may be a terminal that satisfies the CPU type and capacity, GPU type and capacity, RAM type and capacity, etc., which are pre-set based on the CPU type and capacity, GPU type and capacity, RAM type and capacity, etc. of the low-performance terminal (200).

[0037] In addition, the low-performance terminal (200) is registered (or subscribed) as a terminal for performing an intensive learning function on the integrated terminal (400) through linkage with the integrated terminal (400), and registers terminal specifications including a pre-set CPU type and capacity, GPU type and capacity, RAM type and capacity, etc. based on the CPU type and capacity, GPU type and capacity, RAM type and capacity, etc. of the low-performance terminal (200).

[0038] In addition, the low-performance terminal (200), in order to use the functions (or services) provided by the integrated terminal (400) through linkage with the integrated terminal (400), registers a consent form for the use of personal information, a consent form for the collection of personal information, a consent form for the provision of personal information, a consent form for the processing of unique identification information, etc., in the form of electronic documents pre-set in the low-performance terminal (200) to the integrated terminal (400).

[0039] In addition, the low-performance terminal (200) accumulates points provided by the integrated terminal (400) by performing an intensive learning function related to a specific object using raw data provided by the integrated terminal (400). The points accumulated in this way can be used like cash through the shopping mall server, etc.

[0040] Additionally, the low-performance terminal (200) can display (or receive) statistical information regarding the learning results using artificial intelligence performed at the request of the integrated terminal (400) through a dedicated app and / or dedicated website provided by the integrated terminal (400).

[0041] Additionally, the low-performance terminal (200) receives a plurality of sub-raw data allocated to the low-performance terminal (200) provided (or transmitted) from the integrated terminal (400), information about any one of the supervised learning models (or at least one supervised learning model), information about the epoch, etc. At this time, the plurality of sub-raw data may be allocated (or assigned) according to the GPU usage capacity of the low-performance terminal (200) among a plurality of raw data related to a specific object to be learned.

[0042] Additionally, the low-performance terminal (200), according to information regarding any one of the received supervised learning models (or at least one supervised learning model), performs learning (or artificial intelligence / machine learning / deep learning) by using the received (or each assigned to each high-performance terminal) multiple sub-raw data as input values ​​for the corresponding artificial intelligence-based model (or learning model), and generates sub-results related to the corresponding sub-raw data based on the learning results (or artificial intelligence results / machine learning results / deep learning results). Here, the sub-result (or sub-learning result / second sub-result) includes information regarding the corresponding sub-raw data (or including an index, ID, etc. regarding the corresponding raw data), the number of epochs (or the number of epochs), the results of the learning execution, etc. At this time, the low-performance terminal (200) performs a learning function for the corresponding multiple sub-raw data using a CPU (Central Processing Unit) among the various components constituting the low-performance terminal (200).

[0043] In addition, the low-performance terminal (200) transmits the sub-results generated each time an epoch of learning is achieved, identification information of the low-performance terminal (200), etc., to the management terminal (300). Here, the identification information of the low-performance terminal (200) includes Agent Server ID, MDN, mobile IP, mobile MAC, SIM card unique information, serial number, etc.

[0044] In the embodiment of the present invention, it is mainly described that a sub-result related to the sub-raw data is generated by the learning result through the specific supervised learning model selected earlier, but it is not limited thereto. Each of the plurality of low-performance terminals (200) may perform learning in relation to the sub-raw data for each of the plurality of supervised learning models that are pre-set so as to compare the learning results through the plurality of supervised learning models, generate a sub-result for each supervised learning model according to the learning result, and transmit the generated sub-result for each supervised learning model to the management terminal (300).

[0045] The management terminal (300) communicates with the plurality of high-performance terminals (100), the plurality of low-performance terminals (200), the integrated terminal (400), etc. At this time, the management terminal (300) may be a terminal selected by the integrated terminal (400) from among all terminals (100, 200, 300) managed by the integrated terminal (400), and may be selected from among the terminals classified as high-performance terminals (100).

[0046] In addition, the management terminal (300) is registered (or subscribed) as a terminal for performing an intensive learning function on the integrated terminal (400) through linkage with the integrated terminal (400), and registers terminal specifications including a pre-set CPU type and capacity, GPU type and capacity, RAM type and capacity, etc. based on the CPU type and capacity, GPU type and capacity, RAM type and capacity, etc. of the management terminal (300).

[0047] In addition, the management terminal (300), in conjunction with the integrated terminal (400), registers a consent form for the use of personal information, a consent form for the collection of personal information, a consent form for the provision of personal information, a consent form for the processing of unique identification information, etc., in the form of an electronic document pre-set in the management terminal (300) to use the functions (or services) provided by the integrated terminal (400).

[0048] In addition, the management terminal (300) accumulates points provided by the integrated terminal (400) by performing an intensive learning function related to the specific object using raw data provided by the integrated terminal (400). The points accumulated in this way can be used like cash through the shopping mall server, etc.

[0049] Additionally, the management terminal (300) can display (or receive) statistical information regarding the results of learning using artificial intelligence performed on a plurality of high-performance terminals (100) and / or low-performance terminals (200) in relation to the specific object through a dedicated app and / or a dedicated website provided by the integrated terminal (400).

[0050] Additionally, the management terminal (300) receives information regarding a plurality of sub-raw data (or a plurality of sub-raw data for low-performance computation) assigned to each of the plurality of high-performance terminals (100) and / or the plurality of low-performance terminals (200) provided (or transmitted) from the integration terminal (400), information regarding any one selected supervised learning model (or at least one supervised learning model), information regarding the set epoch, the plurality of raw data, etc.

[0051] Additionally, the management terminal (300) interacts with the plurality of high-performance terminals (100) and / or the plurality of low-performance terminals (200) to primarily collect sub-results (or sub-results per raw data / sub-results per sub-raw data) generated according to the learning performed at the plurality of high-performance terminals (100) and / or the plurality of low-performance terminals (200). At this time, the management terminal (300) can collect sub-results in real time from the plurality of high-performance terminals (100) and / or the plurality of low-performance terminals (200) for each epoch, whenever learning is achieved once. At this time, in order to reduce the load on the integrated terminal (400), the management terminal (300) may be linked with the plurality of high-performance terminals (100) and / or the plurality of low-performance terminals (200) to collect sub-results (or sub-results per raw data / sub-results per sub-raw data) generated according to the learning performed at the plurality of high-performance terminals (100) and / or the plurality of low-performance terminals (200).

[0052] Additionally, the management terminal (300) performs additional artificial intelligence-based learning based on the plurality of raw data, the collected sub-results (or sub-results per raw data) related to the plurality of raw data from the plurality of high-performance terminals (100) and / or the plurality of low-performance terminals (200) according to the entire epoch, and then generates additional sub-results (or additional sub-results per raw data) related to the plurality of raw data based on the results of the additional learning.

[0053] That is, the management terminal (300) performs additional learning (or artificial intelligence / machine learning / deep learning) by using the plurality of raw data, the plurality of sub-results for each of the collected raw data, etc., as input values ​​for the corresponding artificial intelligence-based model (or learning model), and generates additional sub-results (or additional sub-results for each raw data) related to the plurality of raw data based on the additional learning results (or additional artificial intelligence results / additional machine learning results / additional deep learning results). At this time, the management terminal (300) may perform the additional learning function for each of the corresponding raw data a number of times equal to a preset ratio (e.g., including 10%, 20%, 30%, etc.) regarding the information on the epoch. Here, the additional sub-result (or additional sub-result for each raw data) includes information on the corresponding raw data (or including an index, ID, etc. for the corresponding raw data), the number of epoch executions (or the number of epoch executions), and the result of the additional learning performance (or additional sub-result).

[0054] In the embodiments of the present invention, it is mainly described that additional sub-results related to the raw data are generated by additional learning results through the specific supervised learning model selected earlier, but the invention is not limited thereto. The management terminal (300) may perform additional learning related to the raw data for each of the aforementioned multiple supervised learning models that are pre-set so as to compare learning results through multiple supervised learning models, and may generate additional sub-results for each supervised learning model according to the additional learning results.

[0055] Additionally, the management terminal (300) merges (or merges) the sub-results collected in the first stage (or sub-results per raw data / sub-results per sub-raw data) and the generated additional sub-results (or additional sub-results per raw data) to generate a final result related to the specific object that is the learning target. Here, the final result includes learning results (e.g., sub-results, additional sub-results, etc.) for each of the plurality of raw data. At this time, the management terminal (300) may reflect (or apply) the sub-results per supervised learning model collected from the plurality of high-performance terminals (100) and / or the plurality of low-performance terminals (200) when generating the final result.

[0056] Additionally, the management terminal (300) transmits the generated final result, identification information of the management terminal (300), etc., to the integrated terminal (300). Here, the identification information of the management terminal (300) includes an Agent Server ID, MDN, mobile IP, mobile MAC, SIM card unique information, serial number, etc.

[0057] The integrated terminal (400) communicates with the plurality of high-performance terminals (100), the plurality of low-performance terminals (200), the management terminal (300), etc.

[0058] Additionally, regarding all terminals managed by the integrated terminal (400), the integrated terminal (400) manages terminals that satisfy the pre-set CPU type and capacity, GPU type and capacity, RAM type and capacity, etc. based on the CPU type and capacity, GPU type and capacity, RAM type and capacity, etc. of the terminals as high-performance terminals (100), and manages terminals that do not satisfy even one condition among the pre-set CPU type and capacity, GPU type and capacity, RAM type and capacity, etc. as low-performance terminals (200). Additionally, the plurality of operable terminals (100, 200) include one or more high-performance terminals (100), one or more low-performance terminals (200), etc.

[0059] In addition, the integrated terminal (400) registers (or manages) pre-configured electronic documents, such as consent forms for the use of personal information, consent forms for the collection of personal information, consent forms for the provision of personal information, and consent forms for the processing of unique identification information, in a DB server (not shown), in relation to the plurality of high-performance terminals (100), the plurality of low-performance terminals (200), the management terminal (300), etc.

[0060] In addition, the integrated terminal (400) provides a bulletin board function for announcements, events, etc.

[0061] In addition, the integrated terminal (400) accumulates (or provides) points that can be used on a shopping mall server, etc., affiliated with one or more high-performance terminals (100), low-performance terminals (200), management terminals (300), etc., which perform learning functions according to artificial intelligence.

[0062] In addition, the integrated terminal (400) collects multiple raw data related to a specific object from multiple collection target servers or terminals (not shown). Here, the raw data is related to a specific object to be learned (e.g., objects, animals, people, etc.), and the data type includes text, images (e.g., still images, videos, etc.), sound (e.g., voice, etc.), and the data type may be structured data or unstructured data.

[0063] In addition, the integrated terminal (400) performs preprocessing (or preprocessing function) on the collected raw data.

[0064] That is, the integrated terminal (400) performs preprocessing to delete data other than information related to preset parameters from the collected raw data in order to manage information to be used for learning.

[0065] In addition, the integrated terminal (400) utilizes multiple raw data (or multiple preprocessed raw data) related to a specific object to be learned, which has been collected in advance, as data for continuous learning (or machine learning / deep learning). Here, the input dataset for learning can perform training and testing functions by dividing the multiple raw data (or multiple preprocessed raw data) into a training set and a test set at a predetermined ratio (e.g., including 7:3, 8:2, etc.). Furthermore, the input dataset for learning includes multiple raw data (or preprocessed raw data) related to the specific object to be collected later. Additionally, the output dataset for learning is a part to be predicted, which includes learning results related to the specific object by learning based on the raw data (or preprocessed raw data) related to the specific object and predicting it later. Here, the training set is a dataset for training a model, and the learning at this time is intended to find optimal parameters. Furthermore, the aforementioned validation set is a dataset used to validate a model that has already been trained. It is designed to select the best model among multiple trained models and, while it participates to some extent in the training process, the validation data itself does not directly participate in the training. Additionally, the aforementioned test set is a dataset used to evaluate the final performance of the model and does not participate in the training process.

[0066] At this time, the integrated terminal (400) generates (or manages) learning data related to the specific object according to the learning through a supervised learning model managed by the integrated terminal (400).

[0067] In addition, the integrated terminal (400) performs class labeling for learning (or supervised learning) on ​​all raw data (or multiple raw data) related to a specific object. Here, the raw data is related to a specific object to be learned (e.g., objects, animals, people, etc.), and the data type includes text, images (e.g., still images, videos, etc.), sound (e.g., voice, etc.), and the data type may be structured data or unstructured data.

[0068] Additionally, the integrated terminal (400) selects (or chooses) one of a plurality of pre-set supervised learning models (or at least one supervised learning model). Here, the plurality of supervised learning models include CNN (Convolutional Neural Network), FCN (Fully Convolutional Network), Autoencoder, RNN (Recurrent Neural Network), GAN (Generative Adversarial Network), Vision Transformer, DNN (Deep Neural Network), RBM (Restricted Boltzmann Machine), DBN (Deep Belief Network), ANN (Artificial Neural Network), etc.

[0069] In addition, the integrated terminal (400) sets (or selects) an epoch representing the number of training iterations for a dataset (or raw data).

[0070] Additionally, the integrated terminal (400) classifies (or sets) the entire raw data into high-performance computing raw data and low-performance computing raw data based on raw data-specific characteristic information. Here, the raw data-specific characteristic information (or raw data characteristic information) includes metadata, data type (e.g., text, image, sound / voice, etc.), data capacity (or data size), frequency characteristics, resolution characteristics, time complexity, jump size, etc.

[0071] That is, if the data type of the raw data is an image, the integrated terminal (400) performs a preset Fourier transform function for each of the class-labeled entire images (or class-labeled raw data).

[0072] Additionally, the integrated terminal (400) performs filtering on each of the Fourier-transformed entire images (or the Fourier-transformed entire class-labeled raw data) through a preset High-Pass Filter (HPF). At this time, the integrated terminal (400) may also perform filtering on each of the Fourier-transformed entire images through a preset Low-Pass Filter (LPF).

[0073] Additionally, the integrated terminal (400) calculates (or computes) an image-specific frequency average value (or an image-specific frequency average value for all pixels after HPF filtering for each raw data) for each of the filtered entire images (or filtered Fourier transformed class-labeled entire raw data). At this time, the integrated terminal (400) may also calculate an image-specific frequency average value (or an image-specific frequency average value for all pixels after LPF filtering for each raw data) for each of the LPF-filtered entire images.

[0074] Additionally, the integrated terminal (400) calculates (or calculates) the overall image frequency average value (or the frequency average value for the entire HPF-filtered raw data) based on the calculated (or calculated) image frequency average value. At this time, the integrated terminal (400) may also calculate the overall image frequency average value (or the frequency average value for the entire LPF-filtered raw data) based on the calculated LPF-filtered image frequency average value.

[0075] Additionally, the integrated terminal (400) checks one or more image frequency average values ​​among the calculated image frequency average values ​​that are greater than or equal to a preset threshold (e.g., 50% of the total image frequency average value) than the calculated total image frequency average value.

[0076] Additionally, the integrated terminal (400) classifies one or more images (or one or more raw data) corresponding to one or more image frequency average values ​​that are greater than or equal to a preset threshold value than the identified calculated (or calculated) overall image frequency average value among the entire image (or the entire raw data) as high-performance computing images (or high-performance computing raw data).

[0077] Additionally, the integrated terminal (400) classifies at least one image (or at least one raw data / the remaining raw data excluding one or more raw data classified as high-performance computing images among the entire image (or the entire raw data)) corresponding to at least one image frequency average value that is less than a preset threshold value than the identified calculated (or calculated) entire image frequency average value as a low-performance computing image (or low-performance computing raw data).

[0078] In this way, the integrated terminal (400) can classify multiple raw data of which the data type is an image into high-performance computing images and low-performance computing images based on frequency characteristics.

[0079] Additionally, if the data type of the raw data is sound (or voice), the integrated terminal (400) performs a preset Fast Fourier Transform (FFT) function for each of the class-labeled whole sound (or the class-labeled whole sound file / the class-labeled raw data). Here, the Fast Fourier Transform can be performed under preset conditions including a preset sampling frequency (e.g., 128 Hz), a sampling size (e.g., 128), an upper frequency band (e.g., 64 Hz), a frequency resolution (e.g., 1 Hz), etc., and the conditions can be set in various ways according to the designer's design.

[0080] Additionally, the integrated terminal (400) calculates (or determines) a sound-specific resolution (e.g., sound-specific time resolution, sound-specific frequency resolution, etc.) (or raw data-specific resolution) for each of the Fast Fourier Transformed entire sound (or Fast Fourier Transformed entire class-labeled raw data). Here, the time resolution is the value obtained by dividing the FFT size (or sampling size) by the sampling frequency, and the frequency resolution is the value obtained by dividing 1 / time resolution or the sampling rate by the FFT size.

[0081] Additionally, the integrated terminal (400) calculates (or calculates) an overall sound resolution average value (or an average value for resolution per raw data) (e.g., sound time resolution average value, sound frequency resolution average value, etc.) based on the sound resolution calculated (or calculated).

[0082] Additionally, the integrated terminal (400) identifies one or more sound resolutions among the calculated sound resolutions that are greater than or equal to a preset threshold value (e.g., 20% of the average value of the total sound resolution) than the average value of the total sound resolution calculated.

[0083] Additionally, the integrated terminal (400) classifies one or more sounds (or one or more raw data) corresponding to one or more sound resolutions that are greater than or equal to another preset threshold value than the identified calculated (or calculated) average sound resolution value among the entire sound (or the entire raw data) as high-performance computing sounds (or high-performance computing raw data).

[0084] Additionally, the integrated terminal (400) classifies at least one sound (or at least one raw data / the remaining raw data excluding one or more raw data classified as high-performance computing sound among the entire sound (or the entire raw data)) corresponding to at least one sound resolution that is less than another preset threshold value than the identified calculated (or calculated) average value of the entire sound resolution, as low-performance computing sound (or low-performance computing raw data).

[0085] In this way, the integrated terminal (400) can classify multiple raw data of which the data type is sound (or voice) into sounds for high-performance computation and sounds for low-performance computation based on resolution characteristics (e.g., time resolution, frequency resolution, etc.).

[0086] In addition, if the data type of the raw data is text, the integrated terminal (400) performs a pre-set Boyer-Moore algorithm on each of the class-labeled whole text data (or class-labeled whole text file / class-labeled raw data) to calculate (or determine) the jump size per text, time complexity per text, etc.

[0087] In addition, the integrated terminal (400) calculates (or calculates) the average jump size of the entire text data, the average time complexity of the entire text data, etc., based on the jump size of the text data and the time complexity of the text data calculated (or calculated) above.

[0088] Additionally, the integrated terminal (400) identifies one or more text data jump sizes (or one or more text data time complexes) that are greater than another threshold value (e.g., 50% of the average value of the average value of the average value of the average value of the jump size of the entire text data / the average value of the average value of the entire text data time complexity) among the calculated jump sizes (or time complexes of the entire text data) that is greater than the calculated average value of the jump size of the entire text data (or the calculated average value of the time complexity of the entire text data).

[0089] Additionally, the integrated terminal (400) classifies one or more text data (or one or more raw data) corresponding to one or more text data jump sizes (or one or more text data time complexes) that are greater than another preset threshold than the identified calculated (or calculated) average value of the entire text data jump size (or the average value of the entire text data time complexity) among the entire text data (or the entire raw data) as high-performance computation text data (or high-performance computation raw data).

[0090] Additionally, the integrated terminal (400) classifies at least one text data (or at least one raw data / the remaining raw data excluding one or more raw data classified as high-performance computation text data) among the entire text data (or the entire raw data) that is less than another threshold value pre-set than the identified corresponding calculated (or calculated) entire text data jump size average value (or the average value of the calculated entire text data time complexity) as low-performance computation text data (or low-performance computation raw data).

[0091] In this way, the integrated terminal (400) can classify multiple raw data of which the data type is text (or text data) into high-performance computation text data and low-performance computation text data, respectively, based on time complexity and / or jump size.

[0092] In the embodiments of the present invention, the case in which the data type of a plurality of raw data related to a specific object is one of image, sound, and text is mainly described, but is not limited thereto. In the case where the data type of a plurality of raw data related to a specific object is composed of a mixture of image, sound, text, etc., the integrated terminal (400) may classify the plurality of raw data into high-performance computation raw data and low-performance computation raw data, respectively, based on frequency characteristics, resolution characteristics, time complexity, jump size, etc., according to characteristic information for each raw data.

[0093] Additionally, the integrated terminal (400) selects (or chooses) a plurality of operable terminals (100, 200) among the entire high-performance terminal (100) and low-performance terminal (200) being managed by the integrated terminal (400) based on terminal-specific model learning status information (e.g., information on whether a learning function using a learning model is being performed), terminal-specific CPU usable capacity (or terminal-specific CPU usable ratio / terminal-specific CPU usage rate), terminal-specific GPU usable capacity (or terminal-specific GPU usable ratio / terminal-specific GPU usage rate), etc. Here, the integrated terminal (400) manages, for all terminals managed by the integrated terminal (400), terminals that satisfy the pre-set CPU type and capacity, GPU type and capacity, RAM type and capacity, etc. based on the CPU type and capacity, GPU type and capacity, RAM type and capacity, etc. of the terminals as high-performance terminals (100), and terminals that do not satisfy even one condition among the pre-set CPU type and capacity, GPU type and capacity, RAM type and capacity, etc. are managed as low-performance terminals (200). In addition, the plurality of operable terminals (100, 200) include one or more high-performance terminals (100), one or more low-performance terminals (200), etc.At this time, the integrated terminal (400) may select (or choose) a plurality of operable terminals (100, 200) among the entire high-performance terminal (100) and low-performance terminal (200) being managed, by considering, in addition to the above conditions (e.g., model learning status information per terminal, CPU usable capacity per terminal, GPU usable capacity per terminal, etc.), RAM usable capacity per terminal (or memory usable capacity per terminal / memory usage rate per terminal), the average CPU usable capacity per terminal during a reference period (e.g., the last 1 hour, 3 hours, 6 hours, etc. from the current time), the average GPU usable capacity per terminal during a reference period, the average RAM usable capacity per terminal during a reference period, and one or more process usage information per terminal.

[0094] That is, the integrated terminal (400) selects (or chooses) a plurality of high-performance terminals (100) among all high-performance terminals (100) managed by the integrated terminal (400) that are not currently performing a learning function using a learning model and have a GPU usable capacity greater than or equal to a preset threshold value, and / or a plurality of low-performance terminals (200) among all low-performance terminals (200) managed by the integrated terminal (400) that are not currently performing a learning function using a learning model and have a CPU usable capacity greater than or equal to another preset threshold value, as a plurality of operating terminals (100, 200) for performing a learning function in relation to the plurality of raw data.

[0095] Additionally, the integrated terminal (400) divides (or classifies / sets) one or more raw data classified as high-performance computing raw data and at least one raw data classified as low-performance computing raw data into multiple sub-raw data (or multiple sub-class labeled raw data / multiple high-performance computing sub-raw data / multiple low-performance computing sub-raw data) based on the CPU usable capacity (or CPU usable ratio / CPU usage rate per terminal) and GPU usable capacity (or GPU usable ratio / GPU usage rate per terminal) of the selected (or chosen) multiple operable terminals (e.g., including the multiple high-performance terminals (100), the multiple low-performance terminals (200), etc.) per terminal and the GPU usable capacity (or GPU usable ratio / GPU usage rate per terminal). At this time, the integrated terminal (400) may divide (or classify) the entire raw data (e.g., one or more raw data classified as high-performance computing raw data, at least one raw data classified as low-performance computing raw data, etc.) into a plurality of sub-raw data by considering, in addition to the other conditions (e.g., CPU usable capacity per terminal, GPU usable capacity per terminal, etc.), RAM usable capacity per terminal (or memory usable capacity per terminal / memory usage rate per terminal), the average CPU usable capacity per terminal reference period during a preset reference period (e.g., the last 1 hour, 3 hours, 6 hours, etc. from the current time), the average GPU usable capacity per terminal reference period, the average RAM usable capacity per terminal reference period, and one or more process usage information per terminal.

[0096] That is, the integrated terminal (400) divides (or classifies / sets) one or more raw data previously classified as raw data for high-performance computation into a plurality of sub-raw data (or a plurality of sub-raw data for high-performance computation) corresponding to each of the plurality of high-performance terminals (100), based on the GPU usable capacity for each of the plurality of high-performance terminals (100) selected (or chosen).

[0097] Additionally, the integrated terminal (400) divides (or classifies / sets) at least one raw data previously classified as raw data for low-performance computation into a plurality of sub-raw data (or a plurality of sub-raw data for low-performance computation) corresponding to each of the respective low-performance terminals (200), based on the CPU usable capacity for each of the respective low-performance terminals (200) associated with the respective low-performance terminals (or selected). At this time, if the selected plurality of high-performance terminals (100) can process not only one or more raw data classified as high-performance computing raw data but also at least one raw data classified as low-performance computing raw data (or all raw data) according to the GPU usage capacity of each high-performance terminal associated with the selected plurality of high-performance terminals (100), the integrated terminal (400) may divide (or classify / set) the entire raw data (or one or more raw data classified as high-performance computing raw data and at least one raw data classified as low-performance computing raw data) into a plurality of sub-raw data corresponding to each of the selected plurality of high-performance terminals (100).

[0098] In addition, the integrated terminal (400) assigns (or assigns) the divided (or classified / configured) multiple sub-row data to each of the previously selected (or chosen) multiple operable terminals (100, 200).

[0099] That is, the integrated terminal (400) allocates (or assigns) a plurality of sub-low data (or a plurality of sub-low data for high-performance computation) corresponding to each of the divided (or classified / configured) corresponding high-performance terminals (100) according to the GPU usage capacity of each of the high-performance terminals (100).

[0100] In addition, the integrated terminal (400) allocates (or assigns) a plurality of sub-row data (or a plurality of sub-row data for low-performance computation) corresponding to each of the divided (or classified / configured) corresponding low-performance terminals (200) according to the CPU usable capacity of each of the plurality of low-performance terminals (200).

[0101] In addition, the integrated terminal (400) provides (or transmits) to each of the plurality of operable terminals (100, 200) a plurality of sub-raw data allocated to each of the plurality of operable terminals, information about any one of the selected supervised learning models (or at least one supervised learning model), information about the set epoch, etc.

[0102] That is, the integrated terminal (400) provides (or transmits) to each of the multiple high-performance terminals (100) corresponding to each of the following: multiple sub-raw data (or multiple sub-raw data for high-performance computation) allocated according to the GPU usable capacity of each of the multiple high-performance terminals (100); information about any one selected supervised learning model (or at least one supervised learning model); information about the set epoch, etc.

[0103] Additionally, the integrated terminal (400) provides (or transmits) to each of the multiple low-performance terminals (200) corresponding to each of the following: multiple sub-raw data (or multiple sub-raw data for low-performance computation) allocated according to the CPU usable capacity of each of the multiple low-performance terminals (200); information about any one selected supervised learning model (or at least one supervised learning model); information about the set epoch, etc.

[0104] Additionally, the integrated terminal (400) provides (or transmits) to the management terminal (300) information regarding a plurality of sub-raw data (or a plurality of sub-raw data for low-performance computation) assigned to each of the plurality of high-performance terminals (100) and / or the plurality of low-performance terminals (200), information regarding any one of the selected supervised learning models (or at least one supervised learning model), information regarding the set epoch, the plurality of raw data, etc.

[0105] In the embodiments of the present invention, it is mainly described that each of the plurality of raw data is assigned to one high-performance terminal (100) or one low-performance terminal (200) for individual learning, but it is not limited thereto, and the integrated terminal (400) may assign the specific raw data included in the plurality of raw data to two or more high-performance terminals (100) and / or low-performance terminals (200) so that learning can be performed on the specific raw data at a plurality of operable terminals (100, 200).

[0106] Accordingly, for the multiple raw data, a learning function may be performed on the same raw data at two or more high-performance terminals (100) and / or low-performance terminals (200).

[0107] At this time, among the plurality of high-performance terminals (100) and / or the plurality of low-performance terminals (200) that are performing a learning function, if there exists a specific high-performance terminal (100) or a specific low-performance terminal (200) in which the state of increased GPU usage of the high-performance terminal (100) (or the state in which the said GPU usage exceeds a preset threshold value) and / or the state of increased CPU usage of the low-performance terminal (200) (or the state in which the said CPU usage exceeds a preset threshold value) persists for a preset time (e.g., including 10 minutes, 30 minutes, 1 hour, etc.), the integrated terminal (400) (or the management terminal (300)) selects another terminal (100, 200) from among the entire terminals (100, 200) that is not performing a learning function and is in an idle state where the available GPU capacity is greater than or equal to the preset threshold value or the available CPU capacity is greater than or equal to the other preset threshold value as a new learning terminal (100, 200), and the specific high-performance terminal (100) or the specific low-performance Controls to terminate the learning function being performed at the terminal (200), and provides to the newly selected learning terminal (100, 200) replacing the specific high-performance terminal (100) or the specific low-performance terminal (200) sub-results up to the corresponding point in time at the specific high-performance terminal (100) or the specific low-performance terminal (200), a plurality of sub-raw data assigned to the specific high-performance terminal (100) or the specific low-performance terminal (200), information about any one of the supervised learning models (or at least one supervised learning model), information about the epoch, etc.

[0108] In addition, the high-performance terminal (100) or low-performance terminal (200) corresponding to the newly selected learning terminal (100, 200) receives sub-results up to a corresponding point in time at the specific high-performance terminal (100) or specific low-performance terminal (200) provided by the integrated terminal (400) (or the management terminal (300)), a plurality of sub-raw data assigned to the specific high-performance terminal (100) or specific low-performance terminal (200), information about any one of the supervised learning models (or at least one supervised learning model), information about the epoch, etc.

[0109] Additionally, the newly selected high-performance terminal (100) or low-performance terminal (200) performs artificial intelligence-based learning for the remaining number of times of the corresponding epoch based on the sub-results up to the corresponding point in time at the received specific high-performance terminal (100) or specific low-performance terminal (200), a plurality of sub-raw data assigned to the specific high-performance terminal (100) or specific low-performance terminal (200), information about any one of the supervised learning models (or at least one supervised learning model), information about the epoch, etc., to generate a sub-result (or sub-learning result / sub-result per operable terminal) which is the learning result, and whenever a sub-result is generated by the learning performance according to the execution of the corresponding epoch, the sub-result is transmitted to the management terminal (300).

[0110] In this way, the integrated terminal (400) (or the management terminal (300)) monitors the status of a plurality of high-performance terminals (100) and / or a plurality of low-performance terminals (200) that are performing a learning function in real time, and if the operating status of a specific high-performance terminal (100) or a specific low-performance terminal (200) does not meet a preset standard condition according to the monitoring result, it selects a new high-performance terminal (100) or a new low-performance terminal (200) that is idle, controls the stopping of learning performance in the specific high-performance terminal (100) or the specific low-performance terminal (200) that does not meet the standard condition, and can be configured to continue performing the learning function through the newly selected new high-performance terminal (100) or new low-performance terminal (200).

[0111] In addition, the integrated terminal (400) receives the final result transmitted from the management terminal (300), identification information of the management terminal (300), etc.

[0112] In addition, the integrated terminal (400) performs a verification function for the final result based on the received final result, a pre-set verification data set corresponding to a specific object related to the plurality of raw data, etc.

[0113] In addition, the integrated terminal (400) performs a test function on the supervised learning model based on the received final result, a preset test set corresponding to a specific object related to the plurality of raw data, etc.

[0114] Additionally, the integrated terminal (400) generates training data including verification eigenvalues, absolute values, minimum values, maximum values, etc., for the entire raw data related to the specific object based on the received final result, the result of the performed verification function, the result of the performed test function, etc. Here, the integrated terminal (400) manages a vector value (or vector) (for example, a list or array having numbers that the terminal can recognize as components) related to the raw data. Also, when the square matrix A of the raw data is linearly transformed, a non-zero vector in which the transformation result by the linearly transformed A is a constant multiple of itself is called an eigenvector, and this constant multiple value is called the verification eigenvalue. Also, the absolute value represents the absolute value of each data component within each vector, the minimum value represents the minimum value of the entire vector data, and the maximum value represents the maximum value of the entire vector data.

[0115] In addition, the integrated terminal (400) manages (or stores) the generated training data by mapping (or matching / linking) it with the corresponding multiple raw data, the final result related to the corresponding multiple raw data, etc.

[0116] In the embodiments of the present invention, the management terminal (300) and the integrated terminal (400) are described separately according to the functions they perform, but are not limited thereto. The management terminal (300) and the integrated terminal (400) may be configured as a single terminal, i.e., the integrated terminal (400), and the integrated terminal (400) may be configured to perform the functions of the management terminal (300).

[0117] In this way, based on the model learning status information for each terminal, the available CPU capacity for each terminal, the available GPU capacity for each terminal, and the characteristic information for each raw data, multiple operable terminals are selected from among all terminals and the entire raw data is divided into multiple sub-raw data. Then, the multiple sub-raw data is assigned to each of the selected operable terminals, and AI-based learning is performed based on the sub-raw data assigned to each of the operable terminals to generate sub-results. The management terminal first collects multiple sub-results based on the learning results from the multiple operable terminals, and performs AI-based learning independently based on the multiple sub-results collected first and the multiple raw data to generate additional sub-results. The multiple sub-results collected first and the generated additional sub-results are merged to generate and transmit a final result. Finally, the integrated terminal performs verification and test functions based on the final result transmitted from the management terminal to generate training data including verification eigenvalues, absolute values, minimum values, maximum values, etc., for the entire raw data.

[0118] Hereinafter, a control method for an artificial intelligence system using focused learning according to the present invention will be described in detail with reference to FIGS. 1 to 3.

[0119] FIGS. 2 and 3 are flowcharts illustrating a control method of an artificial intelligence system using focused learning according to an embodiment of the present invention.

[0120] First, the integrated terminal (400) performs class labeling for learning (or supervised learning) on ​​all raw data (or multiple raw data) related to a specific object. Here, the raw data is related to a specific object to be learned (e.g., objects, animals, people, etc.), and the data type includes text, images (e.g., still images, videos, etc.), sound (e.g., voice, etc.), and the data type may be structured data or unstructured data.

[0121] Additionally, the integrated terminal (400) selects (or chooses) one of a plurality of pre-set supervised learning models (or at least one supervised learning model). Here, the plurality of supervised learning models include CNN (Convolutional Neural Network), FCN (Fully Convolutional Network), Autoencoder, RNN (Recurrent Neural Network), GAN (Generative Adversarial Network), Vision Transformer, DNN (Deep Neural Network), RBM (Restricted Boltzmann Machine), DBN (Deep Belief Network), ANN (Artificial Neural Network), etc.

[0122] Additionally, the integrated terminal (400) sets (or selects) an epoch representing the number of training iterations for a data set (or raw data).

[0123] For example, the integrated terminal (400) performs class labeling for each image of the first to second 2000 images, which are raw data related to the automobile to be learned, selects a CNN model among the plurality of supervised learning models, and sets the epoch to 100 times (S210).

[0124] Subsequently, the integrated terminal (400) classifies (or sets) the entire raw data into high-performance computing raw data and low-performance computing raw data based on raw data-specific characteristic information. Here, the raw data-specific characteristic information (or raw data characteristic information) includes metadata, data type (e.g., text, image, sound / voice, etc.), data capacity (or data size), frequency characteristics, resolution characteristics, time complexity, jump size, etc.

[0125] That is, if the data type of the raw data is an image, the integrated terminal (400) performs a preset Fourier transform function for each of the class-labeled entire images (or class-labeled raw data).

[0126] Additionally, the integrated terminal (400) performs filtering on each of the Fourier-transformed entire images (or the Fourier-transformed entire class-labeled raw data) through a preset High-Pass Filter (HPF). At this time, the integrated terminal (400) may also perform filtering on each of the Fourier-transformed entire images through a preset Low-Pass Filter (LPF).

[0127] Additionally, the integrated terminal (400) calculates (or computes) an image-specific frequency average value (or an image-specific frequency average value for all pixels after HPF filtering for each raw data) for each of the filtered entire images (or filtered Fourier transformed class-labeled entire raw data). At this time, the integrated terminal (400) may also calculate an image-specific frequency average value (or an image-specific frequency average value for all pixels after LPF filtering for each raw data) for each of the LPF-filtered entire images.

[0128] Additionally, the integrated terminal (400) calculates (or calculates) the overall image frequency average value (or the frequency average value for the entire HPF-filtered raw data) based on the calculated (or calculated) image frequency average value. At this time, the integrated terminal (400) may also calculate the overall image frequency average value (or the frequency average value for the entire LPF-filtered raw data) based on the calculated LPF-filtered image frequency average value.

[0129] Additionally, the integrated terminal (400) checks one or more image frequency average values ​​among the calculated image frequency average values ​​that are greater than or equal to a preset threshold (e.g., 50% of the total image frequency average value) than the calculated total image frequency average value.

[0130] Additionally, the integrated terminal (400) classifies one or more images (or one or more raw data) corresponding to one or more image frequency average values ​​that are greater than or equal to a preset threshold value than the identified calculated (or calculated) overall image frequency average value among the entire image (or the entire raw data) as high-performance computing images (or high-performance computing raw data).

[0131] Additionally, the integrated terminal (400) classifies at least one image (or at least one raw data / the remaining raw data excluding one or more raw data classified as high-performance computing images among the entire image (or the entire raw data)) corresponding to at least one image frequency average value that is less than a preset threshold value than the identified calculated (or calculated) entire image frequency average value as a low-performance computing image (or low-performance computing raw data).

[0132] In this way, the integrated terminal (400) can classify multiple raw data of which the data type is an image into high-performance computing images and low-performance computing images based on frequency characteristics.

[0133] Additionally, if the data type of the raw data is sound (or voice), the integrated terminal (400) performs a preset Fast Fourier Transform (FFT) function for each of the class-labeled whole sound (or the class-labeled whole sound file / the class-labeled raw data). Here, the Fast Fourier Transform can be performed under preset conditions including a preset sampling frequency (e.g., 128 Hz), a sampling size (e.g., 128), an upper frequency band (e.g., 64 Hz), a frequency resolution (e.g., 1 Hz), etc., and the conditions can be set in various ways according to the designer's design.

[0134] Additionally, the integrated terminal (400) calculates (or calculates) sound-specific resolution (e.g., sound-specific time resolution, sound-specific frequency resolution, etc.) (or raw data-specific resolution) for each of the Fast Fourier Transformed entire sound (or Fast Fourier Transformed entire class-labeled raw data).

[0135] Additionally, the integrated terminal (400) calculates (or calculates) an overall sound resolution average value (or an average value for resolution per raw data) (e.g., sound time resolution average value, sound frequency resolution average value, etc.) based on the sound resolution calculated (or calculated).

[0136] Additionally, the integrated terminal (400) identifies one or more sound resolutions among the calculated sound resolutions that are greater than or equal to a preset threshold value (e.g., 20% of the average value of the total sound resolution) than the average value of the total sound resolution calculated.

[0137] Additionally, the integrated terminal (400) classifies one or more sounds (or one or more raw data) corresponding to one or more sound resolutions that are greater than or equal to another preset threshold value than the identified calculated (or calculated) average sound resolution value among the entire sound (or the entire raw data) as high-performance computing sounds (or high-performance computing raw data).

[0138] Additionally, the integrated terminal (400) classifies at least one sound (or at least one raw data / the remaining raw data excluding one or more raw data classified as high-performance computing sound among the entire sound (or the entire raw data)) corresponding to at least one sound resolution that is less than another preset threshold value than the identified calculated (or calculated) average value of the entire sound resolution, as low-performance computing sound (or low-performance computing raw data).

[0139] In this way, the integrated terminal (400) can classify multiple raw data of which the data type is sound (or voice) into sounds for high-performance computation and sounds for low-performance computation based on resolution characteristics (e.g., time resolution, frequency resolution, etc.).

[0140] In addition, if the data type of the raw data is text, the integrated terminal (400) performs a pre-set Boyer-Moore algorithm on each of the class-labeled whole text data (or class-labeled whole text file / class-labeled raw data) to calculate (or determine) the jump size per text, time complexity per text, etc.

[0141] In addition, the integrated terminal (400) calculates (or calculates) the average jump size of the entire text data, the average time complexity of the entire text data, etc., based on the jump size of the text data and the time complexity of the text data calculated (or calculated) above.

[0142] Additionally, the integrated terminal (400) identifies one or more text data jump sizes (or one or more text data time complexes) that are greater than another threshold value (e.g., 50% of the average value of the average value of the average value of the average value of the jump size of the entire text data / the average value of the average value of the entire text data time complexity) among the calculated jump sizes (or time complexes of the entire text data) that is greater than the calculated average value of the jump size of the entire text data (or the calculated average value of the time complexity of the entire text data).

[0143] Additionally, the integrated terminal (400) classifies one or more text data (or one or more raw data) corresponding to one or more text data jump sizes (or one or more text data time complexes) that are greater than another preset threshold than the identified calculated (or calculated) average value of the entire text data jump size (or the average value of the entire text data time complexity) among the entire text data (or the entire raw data) as high-performance computation text data (or high-performance computation raw data).

[0144] Additionally, the integrated terminal (400) classifies at least one text data (or at least one raw data / the remaining raw data excluding one or more raw data classified as high-performance computation text data) among the entire text data (or the entire raw data) that is less than another threshold value pre-set than the identified corresponding calculated (or calculated) entire text data jump size average value (or the average value of the calculated entire text data time complexity) as low-performance computation text data (or low-performance computation raw data).

[0145] In this way, the integrated terminal (400) can classify multiple raw data of which the data type is text (or text data) into high-performance computation text data and low-performance computation text data, respectively, based on time complexity and / or jump size.

[0146] For example, the integrated terminal (400) performs a preset Fourier transform on each of the class-labeled images from the first to the second 2000.

[0147] In addition, the integrated terminal (400) performs a filtering function through the high-pass filter for each of the first to second images that have been transformed by the Fourier transformation.

[0148] In addition, the integrated terminal (400) calculates the first image frequency average value to the second image frequency average value for the first image to the second image frequency average value for the first image to the second image frequency average value filtered through the high-pass filter.

[0149] In addition, the integrated terminal (400) calculates the overall image frequency average value based on the calculated first image frequency average value to the second 2000 image frequency average value.

[0150] Additionally, the integrated terminal (400) checks the first image frequency average value to the 1900th image frequency average value, which is 50% or more of the calculated total image frequency average value, among the first image frequency average value to the 2000th image frequency average value.

[0151] Additionally, the integrated terminal (400) classifies the first image corresponding to the frequency average value of the first image to the 1900 image corresponding to the frequency average value of the 1900 image among the first image to the 2000 images as high-performance computing images.

[0152] Additionally, the integrated terminal (400) classifies the 1901st to 2000th images, which are the remainder among the 1st to 2000th images excluding the 1st to 1900th images classified as high-performance computing images, as low-performance computing images (S220).

[0153] Subsequently, the integrated terminal (400) selects (or chooses) a plurality of operable terminals (100, 200) among the entire high-performance terminal (100) and low-performance terminal (200) being managed by the integrated terminal (400) based on terminal-specific model learning status information (e.g., information on whether a learning function using a learning model is being performed), terminal-specific CPU usable capacity (or terminal-specific CPU usable ratio / terminal-specific CPU usage rate), terminal-specific GPU usable capacity (or terminal-specific GPU usable ratio / terminal-specific GPU usage rate), etc. Here, the integrated terminal (400) manages, for all terminals managed by the integrated terminal (400), terminals that satisfy the pre-set CPU type and capacity, GPU type and capacity, RAM type and capacity, etc. based on the CPU type and capacity, GPU type and capacity, RAM type and capacity, etc. of the terminals as high-performance terminals (100), and terminals that do not satisfy even one condition among the pre-set CPU type and capacity, GPU type and capacity, RAM type and capacity, etc. are managed as low-performance terminals (200). In addition, the plurality of operable terminals (100, 200) include one or more high-performance terminals (100), one or more low-performance terminals (200), etc.

[0154] That is, the integrated terminal (400) selects (or chooses) a plurality of high-performance terminals (100) among all high-performance terminals (100) managed by the integrated terminal (400) that are not currently performing a learning function using a learning model and have a GPU usable capacity greater than or equal to a preset threshold value, and / or a plurality of low-performance terminals (200) among all low-performance terminals (200) managed by the integrated terminal (400) that are not currently performing a learning function using a learning model and have a CPU usable capacity greater than or equal to another preset threshold value, as a plurality of operating terminals (100, 200) for performing a learning function in relation to the plurality of raw data.

[0155] For example, the integrated terminal (400) selects the first to ninth high-performance terminals (100) from among all high-performance terminals (100) managed by the integrated terminal (400) that are not currently performing a learning function using a learning model and have a GPU usable capacity greater than or equal to the preset threshold value.

[0156] Additionally, the integrated terminal (400) selects a first low-performance terminal (200) to a second low-performance terminal (200) from among all low-performance terminals (200) managed by the integrated terminal (400) that are not currently performing a learning function using a learning model and whose CPU usable capacity is greater than or equal to another preset standard value (S230).

[0157] Subsequently, the integrated terminal (400) divides (or classifies / sets) one or more raw data classified as high-performance computing raw data and at least one raw data classified as low-performance computing raw data into multiple sub-raw data (or multiple sub-class labeled raw data / multiple high-performance computing sub-raw data / multiple low-performance computing sub-raw data) based on the CPU usable capacity (or CPU usable ratio / CPU usage rate per terminal) and GPU usable capacity (or GPU usable ratio / GPU usage rate per terminal) of the selected (or chosen) multiple operable terminals (e.g., the multiple high-performance terminals (100), the multiple low-performance terminals (200), etc.) per terminal and the GPU usable capacity (or GPU usable ratio / GPU usage rate per terminal).

[0158] That is, the integrated terminal (400) divides (or classifies / sets) one or more raw data previously classified as raw data for high-performance computation into a plurality of sub-raw data (or a plurality of sub-raw data for high-performance computation) corresponding to each of the plurality of high-performance terminals (100), based on the GPU usable capacity for each of the plurality of high-performance terminals (100) selected (or chosen).

[0159] Additionally, the integrated terminal (400) divides (or classifies / sets) at least one raw data previously classified as raw data for low-performance computation into a plurality of sub-raw data (or a plurality of sub-raw data for low-performance computation) corresponding to each of the respective low-performance terminals (200), based on the CPU usable capacity for each of the respective low-performance terminals (200) associated with the respective low-performance terminals (or selected).

[0160] In addition, the integrated terminal (400) assigns (or assigns) the divided (or classified / configured) multiple sub-row data to each of the previously selected (or chosen) multiple operable terminals (100, 200).

[0161] That is, the integrated terminal (400) allocates (or assigns) a plurality of sub-low data (or a plurality of sub-low data for high-performance computation) corresponding to each of the divided (or classified / configured) corresponding high-performance terminals (100) according to the GPU usage capacity of each of the high-performance terminals (100).

[0162] In addition, the integrated terminal (400) allocates (or assigns) a plurality of sub-row data (or a plurality of sub-row data for low-performance computation) corresponding to each of the divided (or classified / configured) corresponding low-performance terminals (200) according to the CPU usable capacity of each of the plurality of low-performance terminals (200).

[0163] In addition, the integrated terminal (400) provides (or transmits) to each of the plurality of operable terminals (100, 200) a plurality of sub-raw data allocated to each of the plurality of operable terminals, information about any one of the selected supervised learning models (or at least one supervised learning model), information about the set epoch, etc.

[0164] That is, the integrated terminal (400) provides (or transmits) to each of the multiple high-performance terminals (100) corresponding to each of the following: multiple sub-raw data (or multiple sub-raw data for high-performance computation) allocated according to the GPU usable capacity of each of the multiple high-performance terminals (100); information about any one selected supervised learning model (or at least one supervised learning model); information about the set epoch, etc.

[0165] Additionally, the integrated terminal (400) provides (or transmits) to each of the multiple low-performance terminals (200) corresponding to each of the following: multiple sub-raw data (or multiple sub-raw data for low-performance computation) allocated according to the CPU usable capacity of each of the multiple low-performance terminals (200); information about any one selected supervised learning model (or at least one supervised learning model); information about the set epoch, etc.

[0166] Additionally, the integrated terminal (400) provides (or transmits) to the management terminal (300) information regarding a plurality of sub-raw data (or a plurality of sub-raw data for low-performance computation) assigned to each of the plurality of high-performance terminals (100) and / or the plurality of low-performance terminals (200), information regarding any one of the selected supervised learning models (or at least one supervised learning model), information regarding the set epoch, the plurality of raw data, etc.

[0167] For example, the integrated terminal (400) comprises, based on the GPU usable capacity of the first high-performance terminal to the GPU usable capacity of the ninth high-performance terminal related to the selected first high-performance terminal to ninth high-performance terminal, images 1 to 1900 classified as images for high-performance computation, images 1 to 250 corresponding to the first high-performance terminal, images 251 to 500 corresponding to the second high-performance terminal, images 501 to 700 corresponding to the third high-performance terminal, images 701 to 900 corresponding to the fourth high-performance terminal, images 901 to 1100 corresponding to the fifth high-performance terminal, images 1101 to 1300 corresponding to the sixth high-performance terminal, images 1301 to 1500 corresponding to the seventh high-performance terminal, images 1501 to 800 corresponding to the eighth high-performance terminal 1700 images are each divided into 1701 to 1900 images corresponding to the 9th high-performance terminal.

[0168] Additionally, the integrated terminal (400) divides the 1901 image to the 2000 image classified for low-performance computation into the 1901 image to the 1950 image corresponding to the 1901 image and the 1951 image to the 2000 image corresponding to the 2000 image corresponding to the 2000 image, respectively, based on the CPU usable capacity of the 1901 low-performance terminal and the CPU usable capacity of the 2000 low-performance terminal.

[0169] Additionally, the integrated terminal (400) assigns the divided images 1 through 250 to the first high-performance terminal, assigns the divided images 251 through 500 to the second high-performance terminal, assigns the divided images 501 through 700 to the third high-performance terminal, assigns the divided images 701 through 900 to the fourth high-performance terminal, assigns the divided images 901 through 1100 to the fifth high-performance terminal, assigns the divided images 1101 through 1300 to the sixth high-performance terminal, assigns the divided images 1301 through 1500 to the seventh high-performance terminal, assigns the divided images 1501 through 1700 to the eighth high-performance terminal, and assigns the divided images 1701 through 1900 to the ninth Allocate to high-performance terminals.

[0170] Additionally, the integrated terminal (400) assigns the divided images 1901 to 1950 to the first low-performance terminal and assigns the divided images 1951 to 2000 to the second low-performance terminal.

[0171] Additionally, the integrated terminal (400) transmits images 1 to 250 assigned to the first high-performance terminal, information regarding the selected CNN model, information regarding the set 100 epochs, etc. to the first high-performance terminal, images 251 to 500 assigned to the second high-performance terminal, information regarding the selected CNN model, information regarding the set 100 epochs, etc. to the second high-performance terminal, images 501 to 700 assigned to the third high-performance terminal, information regarding the selected CNN model, information regarding the set 100 epochs, etc. to the third high-performance terminal, images 701 to 900 assigned to the fourth high-performance terminal, information regarding the selected CNN model, information regarding the set 100 epochs, etc. to the fourth high-performance terminal, and the [unclear] assigned to the fifth high-performance terminal Images 901 to 1100, information regarding the selected CNN model, information regarding the set 100 epochs, etc. are transmitted to the 5th high-performance terminal; images 1101 to 1300 assigned to the 6th high-performance terminal, information regarding the selected CNN model, information regarding the set 100 epochs, etc. are transmitted to the 6th high-performance terminal; images 1301 to 1500 assigned to the 7th high-performance terminal, information regarding the selected CNN model, information regarding the set 100 epochs, etc. are transmitted to the 7th high-performance terminal; images 1501 to 1700 assigned to the 8th high-performance terminal, information regarding the selected CNN model, information regarding the set 100 epochs, etc. are transmitted to the 8th high-performance terminal; images 1701 to 1900 assigned to the 9th high-performance terminal, Information regarding the above-mentioned selected CNN model,Information regarding the above-set 100 epochs, etc. is transmitted to the 9th high-performance terminal, images 1901 to 1950 assigned to the 1st low-performance terminal, information regarding the selected CNN model, information regarding the above-set 100 epochs, etc. are transmitted to the 1st low-performance terminal, and images 1951 to 2000 assigned to the 2nd low-performance terminal, information regarding the selected CNN model, information regarding the above-set 100 epochs, etc. are transmitted to the 2nd low-performance terminal.

[0172] In addition, the integrated terminal (400) transmits information about images assigned to the first to ninth high-performance terminals and the first to second low-performance terminals, information about the selected CNN model, information about the set 100 epochs, the first to 2000 images, etc., to the management terminal (300) (S240).

[0173] Subsequently, the plurality of high-performance terminals (100) and / or the plurality of low-performance terminals (200) receive a plurality of sub-raw data allocated to each terminal provided (or transmitted) from the integrated terminal (400), information about any one of the supervised learning models (or at least one supervised learning model), information about the epoch, etc.

[0174] Additionally, the plurality of high-performance terminals (100) and / or the plurality of low-performance terminals (200) perform artificial intelligence-based learning based on the plurality of sub-raw data assigned to each of the received terminals, information about any one of the supervised learning models (or at least one supervised learning model), information about the epoch, etc., to generate a sub-result (or sub-result per raw data / sub-result per sub-raw data) which is the result of the learning, and whenever a sub-result is generated by the learning performed according to the execution of the epoch, the sub-result is transmitted to the management terminal (300).

[0175] That is, each of the plurality of high-performance terminals (100) receives a plurality of sub-raw data allocated to each terminal provided (or transmitted) from the integrated terminal (400), information about any one of the supervised learning models (or at least one supervised learning model), information about the epoch, etc.

[0176] Additionally, each of the plurality of high-performance terminals (100) performs learning (or artificial intelligence / machine learning / deep learning) by using the received (or each assigned to each high-performance terminal) plurality of sub-raw data as input values ​​for the corresponding artificial intelligence-based model (or learning model) according to information regarding any one of the received supervised learning models (or at least one supervised learning model), and generates a sub-result related to the corresponding sub-raw data based on the learning result (or artificial intelligence result / machine learning result / deep learning result). Here, the sub-result (or sub-learning result / first sub-result) includes information regarding the corresponding sub-raw data (or including an index, ID, etc. regarding the corresponding raw data), the number of epochs (or the number of epochs), the result of the learning execution, etc.

[0177] In addition, each of the plurality of high-performance terminals (100) transmits the sub-result generated each time one epoch of learning is achieved, identification information of the corresponding high-performance terminal (100), etc., to the management terminal (300). Here, the identification information of the high-performance terminal (100) includes Agent Server ID, MDN, mobile IP, mobile MAC, SIM card unique information, serial number, etc.

[0178] Additionally, each of the plurality of low-performance terminals (200) receives a plurality of sub-raw data allocated to each terminal provided (or transmitted) from the integrated terminal (400), information about any one of the supervised learning models (or at least one supervised learning model), information about the epoch, etc.

[0179] Additionally, each of the plurality of low-performance terminals (200) performs learning (or artificial intelligence / machine learning / deep learning) by using the received (or each assigned to each low-performance terminal) plurality of sub-raw data as input values ​​for the corresponding artificial intelligence-based model (or learning model) according to information regarding any one of the received supervised learning models (or at least one supervised learning model), and generates a sub-result related to the corresponding sub-raw data based on the learning result (or artificial intelligence result / machine learning result / deep learning result). Here, the sub-result (or sub-learning result / second sub-result) includes information regarding the corresponding sub-raw data (or including an index, ID, etc. regarding the corresponding raw data), the number of epochs (or the number of epochs), the result of the learning execution, etc.

[0180] In addition, each of the plurality of low-performance terminals (200) transmits the sub-result generated each time one epoch of learning is achieved, identification information of the corresponding low-performance terminal (200), etc., to the management terminal (300). Here, the identification information of the low-performance terminal (200) includes Agent Server ID, MDN, mobile IP, mobile MAC, SIM card unique information, serial number, etc.

[0181] For example, the first high-performance terminal receives the first to 250 images assigned to the first high-performance terminal transmitted from the integrated terminal (400), information about the selected CNN model, information about the set 100 epochs, etc.

[0182] In addition, the first high-performance terminal performs learning by using the first to 250 images, etc. as input values ​​for the CNN model based on the received first to 250 images, information about the received CNN model, information about the received 100 epochs, etc., and generates a 1-1 sub-learning result (e.g., a 1st learning result for the first image), a 1-2 sub-learning result (e.g., a 2nd learning result for the first image), ..., a 1-100 sub-learning result (e.g., a 100th learning result for the first image), a 2-1 sub-learning result (e.g., a 1st learning result for the second image), ..., a 250-100 sub-learning result (e.g., a 100th learning result for the second image), etc.

[0183] In addition, the first high-performance terminal transmits the 1-1 sub-learning result, 1-2 sub-learning result, ..., 1-100 sub-learning result, 2-1 sub-learning result, ..., 250-100 sub-learning result, identification information of the first high-performance terminal, etc. generated according to the learning performance to the management terminal (300) in real time (S250).

[0184] Subsequently, the management terminal (300) interacts with the plurality of high-performance terminals (100) and / or the plurality of low-performance terminals (200) to first collect sub-results (or sub-results per raw data / sub-results per sub-raw data) generated according to the learning performed at the plurality of high-performance terminals (100) and / or the plurality of low-performance terminals (200). At this time, the management terminal (300) can collect sub-results in real time from the plurality of high-performance terminals (100) and / or the plurality of low-performance terminals (200) for each epoch, whenever learning is achieved once.

[0185] For example, the management terminal (300) collects in real time the 1-1 sub-learning result, 1-2 sub-learning result, ..., 1-100 sub-learning result, 2-1 sub-learning result, ..., 250-100 sub-learning result, identification information of the 1-1 high-performance terminal, etc. generated according to learning at the 1-1 high-performance terminal (S260).

[0186] Afterward, the management terminal (300) completes the collection of multiple sub-results (or multiple sub-results per raw data) related to the multiple raw data from multiple high-performance terminals (100) and / or multiple low-performance terminals (200) according to the entire epoch, and performs additional artificial intelligence-based learning based on the multiple raw data, the collected multiple sub-results per raw data, information about any one supervised learning model (or at least one supervised learning model), information about the epoch, etc., and generates additional sub-results (or additional sub-results per raw data) related to the multiple raw data based on the additional learning results.

[0187] That is, the management terminal (300) performs additional learning (or artificial intelligence / machine learning / deep learning) by using the plurality of raw data, the plurality of sub-results for each of the collected raw data, etc., as input values ​​for the corresponding artificial intelligence-based model (or learning model), and generates additional sub-results (or additional sub-results for each raw data) related to the plurality of raw data based on the additional learning results (or additional artificial intelligence results / additional machine learning results / additional deep learning results). At this time, the management terminal (300) may perform the additional learning function for each of the corresponding raw data a number of times equal to a preset ratio (e.g., including 10%, 20%, 30%, etc.) regarding the information on the epoch. Here, the additional sub-result (or additional sub-result for each raw data) includes information on the corresponding raw data (or including an index, ID, etc. for the corresponding raw data), the number of epoch executions (or the number of epoch executions), and the result of the additional learning performance (or additional sub-result).

[0188] For example, the management terminal (300) completes the collection in real time of a plurality of sub-learning results according to the learning of each of the first high-performance terminal to the ninth high-performance terminal and the first low-performance terminal to the second low-performance terminal, and then performs additional learning using the first image to the second 2000 image, etc. as input values ​​for the corresponding CNN model over a total of 10 times (e.g., 100 epochs * 10% = 10 times) according to a first ratio (e.g., 10%) pre-set for the 100 epochs, and performs additional learning using the first image to the second 2000 image, etc. as input values ​​for the corresponding CNN model, and the first-1 additional learning result (e.g., the first additional learning result for the first image), the first-2 additional learning result (e.g., the second additional learning result for the first image), ..., the first-10 additional learning result (e.g., the tenth additional learning result for the first image), ..., the second-10 additional learning result (e.g., the A plurality of additional sub-results are generated, including the result of the 10th additional training on 2000 images (S270).

[0189] Subsequently, the management terminal (300) merges (or merges) the sub-results collected in the first stage (or sub-results per raw data / sub-results per sub-raw data) and the generated additional sub-results (or additional sub-results per raw data) to generate a final result related to the specific object that is the learning target. Here, the final result includes learning results (e.g., sub-results, additional sub-results, etc.) for each of the plurality of raw data.

[0190] Additionally, the management terminal (300) transmits the generated final result, identification information of the management terminal (300), etc., to the integrated terminal (300). Here, the identification information of the management terminal (300) includes an Agent Server ID, MDN, mobile IP, mobile MAC, SIM card unique information, serial number, etc.

[0191] For example, the management terminal (300) generates a first final result by merging a plurality of sub-learning results collected from the first to ninth high-performance terminals and the first to second low-performance terminals, respectively, and the generated plurality of additional learning results, and transmits the generated first final result, identification information of the management terminal (300), etc., to the integrated terminal (400) (S280).

[0192] Afterwards, the integrated terminal (400) receives the final result transmitted from the management terminal (300), identification information of the management terminal (300), etc.

[0193] In addition, the integrated terminal (400) performs a verification function for the final result based on the received final result, a pre-set verification data set corresponding to a specific object related to the plurality of raw data, etc.

[0194] In addition, the integrated terminal (400) performs a test function on the supervised learning model based on the received final result, a pre-set test set corresponding to a specific object related to the plurality of raw data, etc.

[0195] Additionally, the integrated terminal (400) generates training data including verification eigenvalues, absolute values, minimum values, maximum values, etc., for the entire raw data related to the specific object based on the received final result, the result of the performed verification function, the result of the performed test function, etc. Here, the integrated terminal (400) manages a vector value (or vector) (for example, a list or array having numbers that the terminal can recognize as components) related to the raw data. Also, when the square matrix A of the raw data is linearly transformed, a non-zero vector in which the transformation result by the linearly transformed A is a constant multiple of itself is called an eigenvector, and this constant multiple value is called the verification eigenvalue. Also, the absolute value represents the absolute value of each data component within each vector, the minimum value represents the minimum value of the entire vector data, and the maximum value represents the maximum value of the entire vector data.

[0196] In addition, the integrated terminal (400) manages (or stores) the generated training data by mapping (or matching / linking) it with the corresponding multiple raw data, the final result related to the corresponding multiple raw data, etc.

[0197] For example, the integrated terminal (400) receives a first final result transmitted from the management terminal (300), identification information of the management terminal (300), etc.

[0198] In addition, the integrated terminal (400) performs a verification function based on the received first final result, a plurality of pre-set verification data sets related to the vehicle, etc., and generates a first verification function result.

[0199] In addition, the integrated terminal (400) performs a test function based on the received first final result, a plurality of pre-set test sets related to the vehicle, etc., and generates a first test function result.

[0200] In addition, the integrated terminal (400) generates first training data based on the first final result, the first verification function result, the first test function result, the first image to the 2000 image, etc., and stores the generated first training data (S290).

[0201] As previously described, an embodiment of the present invention selects a plurality of operable terminals from among all terminals based on terminal-specific model learning status information, terminal-specific CPU available capacity, terminal-specific GPU available capacity, and raw data characteristic information for the entire raw data, divides the entire raw data into a plurality of sub-raw data, assigns the plurality of sub-raw data to each of the selected operable terminals, performs AI-based learning based on the sub-raw data assigned to each of the operable terminals to generate sub-results, collects the plurality of sub-results based on the learning results from the plurality of operable terminals in the first stage at a management terminal, performs AI-based learning independently based on the collected plurality of sub-results and the plurality of raw data to generate additional sub-results, merges the collected plurality of sub-results and the generated additional sub-results to generate and transmit a final result, and performs verification and test functions based on the final result transmitted from the management terminal at an integrated terminal to generate learning data including verification eigenvalues, absolute values, minimum values, maximum values, etc. for the entire raw data, thereby reducing the training time required through intensive learning of the AI-based model and improving the operational efficiency of the entire system It can be improved.

[0202] A person skilled in the art to which the present invention pertains will be able to make modifications and variations to the foregoing without departing from the essential characteristics of the present invention. Accordingly, the embodiments disclosed in the present invention are intended to explain, not limit, the technical concept of the present invention, and the scope of the technical concept of the present invention is not limited by these embodiments. The scope of protection of the present invention shall be interpreted by the claims below, and all technical concepts within an equivalent scope shall be interpreted as being included within the scope of rights of the present invention. Explanation of the symbols

[0203] 10: Artificial intelligence system using focused learning 100: Multiple high-performance terminals 200: Multiple low-performance terminals 300: Management terminal 400: Integrated terminal

Claims

Claim 1 The integrated terminal performs class labeling for learning on all raw data related to a specific object, classifies all raw data into raw data for high-performance computing and raw data for low-performance computing based on characteristic information for each raw data, selects a plurality of operable terminals from among all high-performance terminals and low-performance terminals managed by the integrated terminal, divides one or more raw data classified as high-performance computing raw data and at least one raw data classified as low-performance computing raw data into a plurality of sub-raw data based on the CPU usable capacity or GPU usable capacity of each of the selected plurality of operable terminals, assigns the divided plurality of sub-raw data to the plurality of high-performance terminals and the plurality of low-performance terminals respectively, and provides the plurality of sub-raw data assigned to the plurality of high-performance terminals and the plurality of low-performance terminals respectively, information on a pre-selected supervised learning model, and information on a pre-set epoch to the plurality of high-performance terminals and the plurality of low-performance terminals respectively; An artificial intelligence system using intensive learning comprising a plurality of high-performance terminals and a plurality of low-performance terminals, wherein the system performs artificial intelligence-based learning based on a plurality of sub-raw data provided from the integrated terminal, information about any one of the supervised learning models, and information about the epoch, thereby generating sub-results which are learning results, and transmits the corresponding sub-results to a management terminal whenever a sub-result is generated by the learning performance according to the execution of the epoch. Claim 2 In claim 1, the management terminal interacts with the plurality of high-performance terminals or the plurality of low-performance terminals to collect sub-results generated according to the performance of learning at the plurality of high-performance terminals or the plurality of low-performance terminals, and after the collection of the plurality of sub-results related to the plurality of raw data from the plurality of high-performance terminals and the plurality of low-performance terminals is completed according to the entire epoch, performs additional AI-based learning based on the plurality of raw data, the plurality of sub-results for each of the collected raw data, any one of the supervised learning models, and information regarding the epoch, generates additional sub-results related to the corresponding plurality of raw data based on the additional learning results, merges the collected sub-results and the generated additional sub-results to generate a final result related to the corresponding specific object that is the subject of learning, and transmits the generated final result to the integration terminal; the integration terminal performs a verification function and a test function, respectively, on the final result transmitted from the management terminal, and generates and stores learning data for the entire raw data related to the specific object based on the final result, the result of the performed verification function, and the result of the performed test function. An artificial intelligence system that uses intensive learning. Claim 3 A step of performing class labeling for learning on all raw data related to a specific object by the integrated terminal; a step of classifying all raw data into raw data for high-performance computing and raw data for low-performance computing based on characteristic information for each raw data by the integrated terminal; a step of selecting a plurality of operable terminals from among all high-performance terminals and low-performance terminals managed by the integrated terminal by the integrated terminal; a step of dividing one or more raw data classified as high-performance computing raw data and at least one raw data classified as low-performance computing raw data into a plurality of sub-raw data by the integrated terminal based on the available CPU capacity or available GPU capacity for each terminal of the plurality of high-performance terminals and the plurality of low-performance terminals selected as operable terminals by the integrated terminal; a step of assigning the divided plurality of sub-raw data to each of the plurality of high-performance terminals and the plurality of low-performance terminals selected by the integrated terminal; and information regarding a predetermined supervised learning model for the plurality of sub-raw data assigned to each of the plurality of high-performance terminals and the plurality of low-performance terminals by the integrated terminal. and, a step of providing information regarding a preset epoch to each of the plurality of high-performance terminals and the plurality of low-performance terminals; a step of, by each of the plurality of high-performance terminals and the plurality of low-performance terminals, performing artificial intelligence-based learning based on a plurality of sub-raw data provided from the integrated terminal, information regarding any one of the supervised learning models, and information regarding the epoch, thereby generating a sub-result as a learning result, and transmitting the corresponding sub-result to a management terminal whenever a sub-result is generated by the learning performed according to the execution of the epoch;A step of collecting sub-results generated according to the performance of learning on the plurality of high-performance terminals or the plurality of low-performance terminals by the management terminal in conjunction with the plurality of high-performance terminals or the plurality of low-performance terminals; a step of, by the management terminal, after the collection of a plurality of sub-results related to the plurality of raw data from the plurality of high-performance terminals and the plurality of low-performance terminals according to the entire epoch, performing additional AI-based learning based on the plurality of raw data, the plurality of sub-results for each collected raw data, any one of the supervised learning models, and information regarding the epoch, and generating additional sub-results related to the corresponding plurality of raw data based on the additional learning results; a step of, by the management terminal, merging the collected sub-results and the generated additional sub-results to generate a final result related to the corresponding specific object that is the learning target, and transmitting the generated final result to the integration terminal; and a step of, by the integration terminal, performing a verification function and a test function, respectively, on the final result transmitted from the management terminal. A method for controlling an artificial intelligence system using intensive learning, comprising the step of generating and storing training data for all raw data related to the specific object based on the final result, the result of the performed verification function, and the result of the performed test function, by the integrated terminal. Claim 4 In claim 3, the step of classifying the entire raw data into raw data for high-performance computation and raw data for low-performance computation comprises: a process of performing a preset Fourier transform function for each of the class-labeled entire images when the data type of the raw data is an image; a process of performing filtering through a preset high-pass filter for each of the Fourier transformed entire images; a process of calculating an image-specific frequency average value for each of the filtered entire images; a process of calculating an entire image frequency average value based on the calculated image-specific frequency average value; a process of identifying one or more image frequency average values ​​among the calculated image-specific frequency average values ​​that are greater than or equal to a preset threshold value than the calculated entire image frequency average value; a process of classifying one or more images among the entire images that correspond to one or more image frequency average values ​​among the entire images that are greater than or equal to a preset threshold value than the identified calculated entire image frequency average value as images for high-performance computation; and a process of classifying the remaining images among the entire images, excluding the one or more images classified as images for high-performance computation, as images for low-performance computation. Claim 5 In claim 3, the step of classifying the entire raw data into raw data for high-performance computation and raw data for low-performance computation comprises: a process of performing a preset Fast Fourier Transform function for each of the entire class-labeled sounds when the data type of the raw data is sound; a process of calculating a sound-specific resolution for each of the entire sounds that have undergone Fast Fourier Transform; a process of calculating an average value of the entire sound resolution based on the calculated sound-specific resolution; a process of identifying one or more sound resolutions among the calculated sound-specific resolutions that are greater than or equal to a preset threshold value than the calculated average value of the entire sound resolution; a process of classifying one or more sounds among the entire sounds that correspond to one or more sound resolutions among the entire sounds that are greater than or equal to a preset threshold value than the identified calculated average value of the entire sound resolution as sounds for high-performance computation; and a process of classifying the remaining sounds among the entire sounds, excluding the one or more sounds classified as sounds for high-performance computation, as sounds for low-performance computation. Claim 6 In claim 3, the step of classifying the entire raw data into raw data for high-performance computation and raw data for low-performance computation comprises: a process of calculating a jump size per text or a time complexity per text by performing a preset Boyer-Moore algorithm on each of the class-labeled entire text data when the data type of the raw data is text; a process of calculating an average jump size value for the entire text data or an average time complexity value for the entire text data based on the calculated jump size value per text data or a time complexity per text data; a process of identifying one or more text data jump sizes or one or more text data time complexes among the calculated jump sizes per text data or the calculated time complexes per text data that are greater than or equal to another preset threshold value than the calculated average jump size value for the entire text data or the calculated average time complexity value for the entire text data; and among the entire text data, a corresponding one or more text data jump sizes or one or more text data time complexes that are greater than or equal to another preset threshold value than the identified calculated average jump size value for the entire text data or the calculated average time complexity value for the entire text data. A control method for an artificial intelligence system using intensive learning, characterized by comprising: a process of classifying one or more text data into text data for high-performance computation; and a process of classifying the remaining raw data, excluding one or more raw data classified as text data for high-performance computation from the entire text data, into text data for low-performance computation. Claim 7 In claim 3, the step of selecting a plurality of operable terminals among all high-performance terminals and low-performance terminals managed by the integrated terminal comprises: a process of selecting a plurality of high-performance terminals among all high-performance terminals managed by the integrated terminal as operable terminals, wherein the learning function using a learning model is not being performed, and the available GPU capacity is greater than or equal to a preset threshold value; and a process of selecting a plurality of low-performance terminals among all low-performance terminals managed by the integrated terminal as operable terminals, wherein the learning function using a learning model is not being performed, and the available CPU capacity is greater than or equal to another preset threshold value.

Images