Method for providing artificial intelligence model including plurality of models and system thereof

Abstract

A method for providing an artificial intelligence model including a plurality of models, according to an embodiment of the present invention, is a method by which a computing system including a memory and a processor provides an artificial intelligence model including a plurality of models, the method comprising the steps of: acquiring a plurality of specialized models, which are independently pre-trained artificial intelligence models; acquiring a router (gating network) which is an artificial intelligence module for controlling model tasks; combining the acquired router with a plurality of specialized models; constructing a MoELM which is a mixture of experts (MoE) architecture-based model according to the combination; and providing output data for predetermined input data on the basis of the constructed MoELM.

Description

Method for providing an artificial intelligence model including multiple models and system thereof

[0001] The present invention relates to a method and system for providing an artificial intelligence model comprising multiple models. More specifically, the present invention relates to a method and system for providing a Mixture of Experts (MoE) model based on multiple independently trained expert models.

[0002] Typically, artificial intelligence (AI) is implemented through multiple AI models and deep learning based on them.

[0003] These artificial intelligences are being developed to provide various services by taking into account the user's context (e.g., context, environment, and / or intention).

[0004] However, when trying to process a specific task based on large amounts of data, there is a limitation in that the computational cost and time required are significant.

[0005] Due to this, there are certain limitations to the use of AI models in the on-device environment, which has recently been attracting attention.

[0006] To solve this problem, model architectures such as MoE (Mixture of Experts) have been used.

[0007] Here, MoE can mean the architecture of a machine learning model that solves complex problems by combining multiple expert models.

[0008] Such MoEs may include expert models, which are multiple smaller networks designed to learn different parts and / or different characteristics of a given data and perform data processing operations accordingly, and a gating network that evaluates the performance of each expert model and, based on the performance thereof, determines which expert model is most appropriate to assign a specific task to for a given data.

[0009] Thus, according to the MoE architecture, a gating network that acquires a given input data determines a probabilistic or deterministic task assignment for each expert model, and the selected expert models perform their own tasks and return the results, thereby performing data processing for a specific task.

[0010] By leveraging this MoE, AI models can improve overall efficiency and performance by focusing computational resources on specific parts, such as when handling complex tasks or large data sets.

[0011] However, in the case of conventional MoE, not only does it require a high level of VRAM, but there are also significant challenges to be solved during the fine tuning process.

[0012] In addition, the conventional MoE method is designed to efficiently manage large-sized models, and has limitations in supporting the efficiency of remaining resources that are not activated according to a given task.

[0013] In addition, in the conventional field of technology, most services are provided using AI models implemented universally, which poses the problem of making it difficult to quickly and easily secure AI analysis performance most suitable for a given context.

[0014] One embodiment of the present invention is designed to solve the problems described above, and aims to implement a method and system for providing a MoE (Mixture of Experts) model based on a plurality of independently learned expert models.

[0015] However, the technical problems to be solved by the present invention and embodiments of the present invention are not limited to the technical problems described above, and other technical problems may exist.

[0016] A method for providing an artificial intelligence model including a plurality of models according to an embodiment of the present invention is a method for providing an artificial intelligence model including a plurality of models by a computing system including a memory and a processor, the method comprising: a step of obtaining a plurality of specialized models, which are independently learned artificial intelligence models; a step of obtaining a router (Gating Network), which is an artificial intelligence module that controls a model task; a step of combining the obtained router and the plurality of specialized models; a step of constructing an MoELM, which is a model based on an MoE (Mixture of Experts) architecture according to the combination; and a step of providing output data for predetermined input data based on the constructed MoELM.

[0017] In another aspect, the router performs task assignment for each of the plurality of specialized models based on the input data.

[0018] In another aspect, the step of combining the router and the plurality of specialized models includes the step of detecting an input data type for each of the plurality of specialized models, and the step of performing a combining process in different ways depending on whether the detected input data types are identical.

[0019] In another aspect, the step of performing the above combining process includes a step of performing combining between the router and the plurality of specialized models by using the acquired router as is, if the input data types between the plurality of specialized models are identical to each other.

[0020] In another aspect, the step of performing a combination between the router and the plurality of specialized models using the acquired router as is includes a step of integrating the weight values ​​of the router for each of the plurality of specialized models according to a predetermined operation.

[0021] In another aspect, the step of performing the above combining process includes a step of performing combining between the router and the plurality of expert models by additionally learning the acquired router when the input data types between the plurality of expert models are mutually different.

[0022] In another aspect, the step of performing additional learning on the acquired router and performing a combination between the router and the plurality of expert models includes the step of performing additional learning to expand the input data type of each of the plurality of expert models.

[0023] In another aspect, the step of performing additional learning to expand the input data type of each of the plurality of expert models includes the step of mutually integrating the input data of each of the plurality of expert models according to a predetermined operation while preserving the original form of the input data of each of the plurality of expert models, and the step of performing additional learning to set the input data type according to the integrated input data to the input data type of each of the plurality of expert models.

[0024] In another aspect, the step of performing additional learning on the acquired router and performing a combination between the router and the plurality of expert models includes the step of performing additional learning that changes the input data type of each of the plurality of expert models.

[0025] In another aspect, the step of performing additional learning to change the input data type of each of the plurality of expert models includes the step of transforming the input data of a predetermined first expert model into a weight value and transforming it into the input data type of a predetermined second expert model, and the step of performing additional learning to train and operate the second expert model based on the transformed data value.

[0026] In another aspect, the step of combining the router and the plurality of specialized models includes the step of performing additional learning on the acquired router to learn a predetermined existing MoE architecture and automatically set a new MoE based on the learned existing MoE architecture, and the step of performing combining between the router and the plurality of specialized models based on the additionally learned router.

[0027] Meanwhile, a system for providing an artificial intelligence model including a plurality of models according to an embodiment of the present invention comprises: at least one memory; and at least one processor for reading out at least one application stored in the memory and providing an artificial intelligence model including a plurality of models; wherein instructions of the processor include instructions for performing a step of acquiring a plurality of specialized models, which are independently learned artificial intelligence models; a step of acquiring a router (Gating Network), which is an artificial intelligence module for controlling a model task; a step of combining the acquired router and the plurality of specialized models; a step of constructing an MoELM, which is a MoE (Mixture of Experts) architecture-based model according to the combination; and a step of providing output data for predetermined input data based on the constructed MoELM.

[0028] A method for providing an artificial intelligence model including a plurality of models according to an embodiment of the present invention and a system thereof generate and provide output data according to task processing based on predetermined input data using a combined model constructed in the form of MoE by collecting independently learned single models, thereby having the effect of implementing an MoE mechanism that implements highly efficient data processing without wasting unnecessary resources by utilizing small-sized models optimized for a given task, unlike the existing MoE method that operates while maintaining a large-sized model overall.

[0029] Accordingly, the method for providing an artificial intelligence model including a plurality of models according to one embodiment of the present invention and the system thereof can support various services using an improved MoE model that maintains the advantages of the conventional MoE method (e.g., faster pre-learning speed compared to FFN, faster inference speed compared to models of the same size, and / or improved Instruction Tuning performance, etc.) while minimizing the disadvantages of the conventional MoE method (e.g., the need for a high level of VRAM and / or various challenges in the Fine Tuning process, etc.), thereby directly improving the performance and quality of the corresponding service.

[0030] However, the effects that can be obtained from the present invention are not limited to the effects mentioned above, and other effects that are not mentioned can be clearly understood from the description below.

[0031] FIG. 1 illustrates an example of a block diagram of a computing system that implements a model provision service based on MoE architecture according to one embodiment of the present invention.

[0032] FIG. 2 illustrates an example block diagram of a computing device that implements a model provision service based on MoE architecture according to one embodiment of the present invention.

[0033] FIG. 3 illustrates an example block diagram from another aspect of a computing device implementing a model provision service based on MoE architecture according to one embodiment of the present invention.

[0034] FIG. 4 illustrates an internal block diagram of an AI agent model according to one embodiment of the present invention.

[0035] FIG. 5 illustrates a flowchart for explaining a method for providing an MoE model according to one embodiment of the present invention.

[0036] FIG. 6 illustrates a conceptual diagram for explaining a method for providing an MoE model according to one embodiment of the present invention.

[0037] FIG. 7 illustrates a flowchart for explaining a model specification method based on MoE according to one embodiment of the present invention.

[0038] FIG. 8 illustrates a conceptual diagram for explaining a model specification method based on MoE according to one embodiment of the present invention.

[0039] Figure 9 illustrates an example of specialized model characteristic information according to one embodiment of the present invention.

[0040] FIG. 10 illustrates a flowchart for explaining a method for providing an AI agent based on a MoE application LLM according to one embodiment of the present invention.

[0041] FIG. 11 illustrates a conceptual diagram for explaining a method for providing an AI agent based on a MoE application LLM according to one embodiment of the present invention.

[0042] The present invention is capable of various modifications and embodiments. Therefore, specific embodiments are illustrated in the drawings and described in detail in the detailed description. The effects and features of the present invention, as well as the methods for achieving them, will become clear with reference to the embodiments described in detail below together with the drawings. However, the present invention is not limited to the embodiments disclosed below and can be implemented in various forms. In the following embodiments, the terms "first," "second," etc. are not used in a limiting sense but are used for the purpose of distinguishing one component from another. Furthermore, the singular expression includes the plural expression unless the context clearly indicates otherwise. Furthermore, terms such as "include" or "have" indicate the presence of a feature or component described in the specification, and do not preemptively exclude the possibility that one or more other features or components may be added. Furthermore, in the drawings, the sizes of components may be exaggerated or reduced for convenience of explanation. For example, the size and thickness of each component shown in the drawings are arbitrarily shown for convenience of explanation, and thus the present invention is not necessarily limited to what is shown.

[0043] Hereinafter, embodiments of the present invention will be described in detail with reference to the attached drawings. When describing with reference to the drawings, identical or corresponding components are given the same reference numerals and redundant descriptions thereof will be omitted.

[0044]

[0045] [Example system implementing a model provisioning service based on the MoE architecture]

[0046] Hereinafter, an exemplary system implementing a model provision service based on the MoE architecture that provides a MoE (Mixture of Experts) model based on multiple independently trained expert models is described in detail with reference to the attached drawings.

[0047] FIG. 1 illustrates an example of a block diagram of a computing system that implements a model provision service based on MoE architecture according to one embodiment of the present invention.

[0048] Referring to FIG. 1, a computing system (1000) implementing a model provision service based on the MoE architecture of the present invention includes a user computing device (110), a server computing system (130), and a training computing system (150), and the devices are capable of communicating via a network (170).

[0049] A model provision service based on an MoE architecture that provides an MoE (Mixture of Experts) model based on a plurality of independently learned expert models according to one embodiment of the present invention may be implemented and provided locally by a user computing device (110), implemented and provided in the form of a web service by a server computing system (130) communicating with the user computing device (110), or implemented and provided by a user computing device (110) and a server computing system (130) in connection with each other.

[0050] At this time, in the embodiment, the user computing device (110) and / or the server computing system (130) can train the machine learning model (120 and / or 140) through interaction with a training computing system (150) communicatively connected via a network (170). The training computing system (150) may be separate from the server computing system (130) or may be part of the server computing system (130).

[0051] And at this time, the artificial intelligence model can be 1) trained directly locally by the user computing device (110), 2) trained by the server computing system (130) and the user computing device (110) interacting with each other through a network (170), and 3) trained by a separate training computing system (150) using various training techniques and learning techniques. And the artificial intelligence model trained by the training computing system (150) can be implemented in a manner of being provided / updated by transmitting the model to the user computing device (110) and / or the server computing system (130) through a network (170).

[0052] In some embodiments, the training computing system (150) may be part of a server computing system (130) or part of a user computing device (110).

[0053] The user computing device (110) may include any type of computing device, such as a smart phone, a mobile phone, a digital broadcasting device, a personal digital assistant (PDA), a portable multimedia player (PMP), a desktop, a wearable device, an embedded computing device, and / or a tablet PC.

[0054] Such a user computing device (110) includes at least one processor (111) and memory (112). Here, the processor (111) may be composed of at least one or a plurality of electrically connected processors among a central processing unit (CPU), a graphics processing unit (GPU), application specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, micro-controllers, microprocessors, and / or other electrical units for performing functions.

[0055] The memory (112) may include one or more non-transitory / transitory computer-readable storage media such as RAM, ROM, EEPROM, EPROM, flash memory devices, magnetic disks, etc., and combinations thereof, and may include web storage of a server that performs a memory storage function on the Internet. The memory (112) may store data (113) and commands (114) necessary for the at least one processor (111) to perform functional operations such as training an artificial intelligence model or executing a MoE architecture-based model provision service through an artificial intelligence model.

[0056] In one embodiment, the user computing device (110) may store at least one machine learning model (120).

[0057] In detail, the machine learning model (120) may be various machine learning models such as multiple neural networks (e.g., deep neural networks) or other types of machine learning models including nonlinear models and / or linear models, and may be composed of a combination thereof.

[0058] At this time, the neural network may include at least one of feed-forward neural networks, recurrent neural networks (e.g., long short-term memory recurrent neural networks), convolutional neural networks, and / or other types of neural networks.

[0059] In one embodiment, a user computing device (110) may receive at least one machine learning model (120) from a server computing system (130) via a network (170), store the model in a memory (112), and then execute the stored machine learning model (120) by a processor (111) to perform a model provision service based on an MoE architecture.

[0060] In another embodiment, the server computing system (130) may include at least one machine learning model (140), perform operations through the machine learning model (140), and communicate data related thereto with the user computing device (110) in conjunction with the user computing device (110) to provide a model provision service based on the MoE architecture to the user.

[0061] For example, a user computing device (110) can perform a model provision service based on the MoE architecture in a manner that a server computing system (130) provides output for a user's input using a machine learning model (140) via the web.

[0062] Additionally, the artificial intelligence model may be implemented in such a way that at least some of the machine learning models (120 and / or 140) are executed on a user computing device (110) and the rest are executed on a server computing system (130).

[0063] In addition, the user computing device (110) may include at least one input component (121) that detects a user's input. For example, the user input component (121) may include a touch sensor (e.g., a touch screen and / or a touch pad, etc.) that detects a touch of a user's input medium (e.g., a finger or a stylus), an image sensor that detects a user's motion input, a microphone, a button, a mouse, and / or a keyboard that detects a user's voice input, etc. In addition, the user input component (121) may include an interface and an external controller when receiving an input to an external controller (e.g., a mouse and / or a keyboard, etc.) through an interface.

[0064] The server computing system (130) includes at least one processor (131) and memory (132). Here, the processor (131) may be composed of at least one or a plurality of electrically connected processors among a central processing unit (CPU), a graphics processing unit (GPU), application specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, micro-controllers, microprocessors, and / or other electrical units for performing functions.

[0065] And the memory (132) may include one or more non-transitory / transitory computer-readable storage media such as RAM, ROM, EEPROM, EPROM, flash memory devices, magnetic disks, etc., and combinations thereof. This memory (132) may store data (133) and instructions (134) necessary for the processor (131) to perform functional operations such as training an artificial intelligence model or executing a MoE architecture-based model provision service through an artificial intelligence model.

[0066] In one embodiment, the server computing system (130) may be implemented to include at least one computing device. For example, the server computing system (130) may be implemented to operate multiple computing devices according to a sequential computing architecture, a parallel computing architecture, or a combination thereof. Additionally, the server computing system (130) may include multiple computing devices connected via a network (170).

[0067] Additionally, the server computing system (130) may store at least one machine learning model (140). For example, the server computing system (130) may include a neural network and / or other multi-layer nonlinear model as the machine learning model (140). Exemplary neural networks may include feedforward neural networks, deep neural networks, recurrent neural networks, and convolutional neural networks.

[0068] The training computing system (150) includes at least one processor (151) and memory (152). Here, the processor (151) may be composed of at least one or a plurality of electrically connected processors among a central processing unit (CPU), a graphics processing unit (GPU), application specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, micro-controllers, microprocessors, and / or other electrical units for performing functions.

[0069] And the memory (152) may include one or more non-transitory / transitory computer-readable storage media such as RAM, ROM, EEPROM, EPROM, flash memory devices, magnetic disks, etc., and combinations thereof. This memory (152) may store data (153) and instructions (154) necessary for the processor (151) to perform learning of an artificial intelligence model, etc.

[0070] For example, the training computing system (150) may include a model trainer (160) that trains a machine learning model (120 and / or 140) stored in the user computing device (110) and / or server computing system (130) using various training or learning techniques, such as backpropagation of errors (according to the framework illustrated in FIG. 3).

[0071] For example, such a model trainer (160) may perform updates to one or more parameters of a machine learning model (120 and / or 140) in a backpropagation manner based on a defined loss function.

[0072] In some implementations, performing backward propagation of errors may include performing truncated backpropagation through time. The model trainer (160) may perform a number of generalization techniques (e.g., weight reduction, dropout, and / or knowledge distillation) to improve the generalization ability of the trained machine learning model (120 and / or 140).

[0073] In particular, the model trainer (160) can train the machine learning model (120 and / or 140) based on a series of training data (161). Here, the training data (161) can include data in different formats, such as, for example, images, audio samples, and / or text. Examples of image types that can be used can include video frames, LiDAR point clouds, X-ray images, computed tomography scans, hyperspectral images, and / or various other forms of images.

[0074] Such training data (161) may be provided by a user computing device (110) and / or a server computing system (130). When the training computing device trains a machine learning model (120 and / or 140) on specific data of the user computing device (110), the machine learning model (120 and / or 140) may be characterized as a personalized model.

[0075] And the model trainer (160) includes computer logic utilized to provide the desired function.

[0076] Additionally, the model trainer (160) may be implemented as hardware, firmware, and / or software that controls a general-purpose processor. In one implementation, the model trainer (160) includes a program file stored in a storage device, which may be loaded into a memory (152) and executed by one or more processors (151). In another implementation, the model trainer (160) includes one or more sets of computer-executable data (153) and instructions (154) stored in a tangible computer-readable storage medium, such as a RAM hard disk or an optical or magnetic medium.

[0077] The network (170) includes, but is not limited to, a 3rd Generation Partnership Project (3GPP) network, a Long Term Evolution (LTE) network, a World Interoperability for Microwave Access (WIMAX) network, the Internet, a Local Area Network (LAN), a Wireless Local Area Network (Wireless LAN), a Wide Area Network (WAN), a Personal Area Network (PAN), a Bluetooth network, a satellite broadcasting network, an analog broadcasting network, and / or a Digital Multimedia Broadcasting (DMB) network.

[0078] In general, communication over a network (170) may be performed using any type of wired and / or wireless connection, using various communication protocols (e.g., TCP / IP, HTTP, SMTP, and / or FTP), encodings or formats (e.g., HTML and / or XML), and / or protection schemes (e.g., VPN, Secure HTTP, and / or SSL).

[0079] FIG. 2 illustrates an example block diagram of a computing device that implements a model provision service based on MoE architecture according to one embodiment of the present invention.

[0080] As shown in FIG. 2, the computing device (100) included in the user computing device (110), the server computing system (130), and the training computing system (150) includes a plurality of applications (e.g., Application 1 to Application N). Each application may include a machine learning library and one or more machine learning models. For example, the applications may include an image processing (e.g., Detection, Classification, and / or Segmentation, etc.) application, a text messaging application, an email application, a dictation application, a virtual keyboard application, a browser application, and / or a chat-bot application.

[0081] In an embodiment, the computing device (100) may include a model trainer (160) for training an artificial intelligence model, and may store and operate the trained artificial intelligence model to provide output data according to predetermined input data (e.g., text, voice, image, video, and / or sensing data based on a specific sensor, etc.).

[0082] Each application of the computing device (100) may communicate with a number of other components of the computing device (100), such as, for example, one or more sensors, a context manager, a device state component, and / or additional components. In one embodiment, each application may communicate with each device component using an API (e.g., a public API). In one embodiment, the API used by each application may be specific to that application.

[0083] FIG. 3 illustrates an example block diagram from another aspect of a computing device implementing a model provision service based on MoE architecture according to one embodiment of the present invention.

[0084] Referring to FIG. 3, a computing device (200) includes a plurality of applications (e.g., Application 1 to Application N). Each application may communicate with a central intelligence layer. For example, the applications may include an image processing application, a text messaging application, an email application, a dictation application, a virtual keyboard application, and / or a browser application. In one embodiment, each application may communicate with the central intelligence layer (and models stored therein) using an API (e.g., a common API across all applications).

[0085] The central intelligence layer may include multiple machine learning models. For example, as illustrated in FIG. 3, at least some of the machine learning models may be provided to each application and managed by the central intelligence layer. In other implementations, two or more applications may share a single machine learning model. For example, in some implementations, the central intelligence layer may provide a single model to all applications. In some implementations, the central intelligence layer may be included within the operating system of the computing device (200) or implemented differently.

[0086] The central intelligence layer may communicate with a central device data layer. The central device data layer may be a centralized data repository for the computing device (200). As illustrated in FIG. 3, the central device data layer may communicate with a number of other components of the computing device (200), such as, for example, one or more sensors, a context manager, a device state component, and / or additional components. In some implementations, the central device data layer may communicate with each device component using an API (e.g., a private API).

[0087] The techniques described herein may refer to servers, databases, software applications, and other computer-based systems, as well as actions taken and information transmitted to or from such systems. It will be appreciated that the inherent flexibility of computer-based systems allows for a wide range of possible configurations, combinations, and division of labor and functionality between and among components. For example, the processes described herein may be implemented using a single device or component, or multiple devices or components operating in combination. Databases and applications may be implemented on a single system or in a distributed system across multiple systems. Distributed components may operate sequentially or in parallel.

[0088]

[0089] [Artificial Intelligence Agent Specialization Model (AIAM)]

[0090] In another aspect, the computing system (1000) as described above may include an AI agent specialization model (AIAM) according to an embodiment of the present invention.

[0091] Here, the AI ​​agent specialization model (AIAM) according to the embodiment is an AI agent model that applies the MoE (Mixture of Experts) architecture implemented according to the embodiment of the present invention, and may be an artificial intelligence model that includes a data processing algorithm that can act autonomously in a specific environment, solve a task, and achieve a goal.

[0092] These AI agent specialized models (AIAM) may include data processing algorithms to implement cognitive abilities to collect and interpret data from a given environment, decision mechanisms to determine optimal actions based on collected data, execution abilities to execute the determined actions, and learning abilities to improve actions through experience.

[0093] In an embodiment, an AI agent specialized model (AIAM) may acquire certain input data (e.g., text, voice, image, video, and / or sensing data based on a specific sensor, etc.) and provide output data (e.g., response data for a specific query and / or control signals according to a specific command, etc.) by performing a certain task based on the acquired input data.

[0094] FIG. 4 illustrates an internal block diagram of an AI agent model according to one embodiment of the present invention.

[0095] In detail, referring to FIG. 4, the AI ​​agent model according to the embodiment may include at least one router (RT: Router, Gating Network), an orchestrator (OCT: Orchestrator), a small Large Language Model (sLLM), a normal MoE model (NM: Normal MoE Model), an external model (EM: External Model), and / or a specialized model (SM: Specialized Model).

[0096] At this time, in Fig. 4, the AI ​​agent model is described as including the components described above to prevent blurring of features according to an embodiment of the present invention.

[0097] However, it is obvious that a person skilled in the art can understand that, depending on the embodiment, other general components may be included in addition to the components illustrated in FIG. 4, or some of the components illustrated in FIG. 4 may be omitted.

[0098] In more detail, a router (RT: Router, Gating Network) according to an embodiment of the present invention may be an artificial intelligence module that performs task allocation and / or traffic adjustment for multiple models within the MoE architecture.

[0099] Specifically, the router (RT) can analyze given input data and / or request tasks, etc., to determine which model is most suitable for processing the data.

[0100] At this time, the router (RT) can determine a model optimized for processing given data based on the performance, expertise, and / or previous experience of each model.

[0101] Additionally, the router (RT) can support efficient data processing by distributing a given task to at least one model, taking into account the system load.

[0102] Additionally, the router (RT) can dynamically respond to real-time system changes and adjust tasks assigned to specific models.

[0103] In an embodiment, such a router (RT) may be an artificial intelligence module that selectively determines a model (hereinafter, domain-specific model) that executes data processing operations optimized for a given domain.

[0104] That is, in the embodiment, the router (RT) may be an artificial intelligence module that selects a model (i.e., a domain-specific model) that is judged to perform data processing (e.g., deep learning, etc.) most suitable for a given domain among multiple models included in the AI ​​agent specialized model (AIAM).

[0105] For reference, a domain according to an embodiment may refer to data, rules, terminology, problem definitions, and / or processes that a given AI system uses to perform a specific task.

[0106] In an embodiment, a router (RT) may perform data analysis based on characteristics of certain input data (e.g., user input and / or specific sensing data, etc.) and / or requested tasks, and based on this, determine data processing characteristics optimized for the task, detect a certain model implementing the same, and determine a domain-specific model.

[0107] In other words, the router (RT) according to the embodiment may be an artificial intelligence module that detects a model that can most effectively perform data processing according to a given domain, allocates / distributes the corresponding task processing work, and manages it.

[0108] At this time, the router (RT) according to the embodiment may include a router (RT) previously learned according to a disclosed predetermined algorithm, a router (RT) additionally learned according to the embodiment of the present invention, and / or a router (RT) newly learned in a new manner. A detailed description thereof is described in the [MoE model provision method] described below.

[0109] Additionally, the router (RT) according to the embodiment can further perform additional functional operations disclosed in the [MoE model provision method] below.

[0110] Meanwhile, an orchestrator (OCT) according to an embodiment of the present invention may be an artificial intelligence module that comprehensively controls and manages the overall configuration of an AI agent specialization model (AIAM).

[0111] In detail, in an embodiment, an orchestrator (OCT) can assign various tasks occurring in the entire system to appropriate resources (e.g., a router (RT) and / or a predetermined model, etc.).

[0112] Additionally, the orchestrator (OCT) can manage resources such as available models and hardware resources (e.g., CPU and / or GPU) to ensure efficient use.

[0113] Additionally, the orchestrator (OCT) can monitor the performance of the entire system and adjust specific parameters or optimize network configurations, etc. as needed.

[0114] Additionally, the orchestrator (OCT) can manage the interconnection between multiple routers (RTs) and / or models and control the data flow and processing process.

[0115] That is, in the embodiment, the orchestrator (OCT) performs control and management of the entire system of the AI ​​Agent Specialization Model (AIAM) and can perform the role of a main router (RT) that controls at least one router (RT).

[0116] At this time, the orchestrator (OCT) and router (RT) according to the embodiment can be closely linked to each other to support efficient operation of the MoE system.

[0117] Specifically, the orchestrator (OCT) monitors the performance of the router (RT) as the manager of the entire system and can adjust the strategy of the router (RT) as needed.

[0118] On the other hand, the router (RT) can implement efficient system control by actually performing the allocation of data processing tasks according to the instructions of the orchestrator (OCT) and / or its own algorithm.

[0119] In an embodiment, the orchestrator (OCT) and / or router (RT) as described above may be a master model (P) that can control and manage the remaining components of the overall system and / or AI agent models (i.e., sLLM, general MoE model (NM), external model (EM), and / or expert model (SM), etc.).

[0120] Meanwhile, the sLLM (small Large Language Model) according to an embodiment of the present invention may be an artificial intelligence module implemented as a lightweight version of the LLM (Large Language Model).

[0121] In other words, sLLM may be an artificial intelligence module built to achieve performance similar to large models such as LLM with fewer resources.

[0122] In an embodiment, such sLLM may include a MoE model (in an embodiment, MoELM) based on a combination of multiple expert models (SMs) and routers (RTs) according to an embodiment of the present invention disclosed in [MoE model provision method] described below. In addition, the sLLM may include a MoE model (in an embodiment, DMoE model) based on a domain-specific expert model according to an embodiment of the present invention disclosed in [MoE-based model specification method] described below. Detailed descriptions thereof are described in the above [MoE model provision method] and [MoE-based model specification method].

[0123] In addition, the normal MoE model (NM) according to an embodiment of the present invention may mean a predetermined MoE model implemented according to the disclosed universal method.

[0124] For example, a general MoE model (NM) may include Switch Transformer, Conditional Computation in Neural Networks, Sparse Mixture of Experts, and / or Megatron-LM.

[0125] In addition, an external model (EM) according to an embodiment of the present invention may mean a predetermined artificial intelligence model implemented according to various disclosed algorithms.

[0126] For example, external models (EMs) may include ChatGPT, Gemini, and / or Llama.

[0127] In embodiments, such external models (EMs) may be optionally used as needed to support processing of a given task.

[0128] In addition, a specialized model (SM) according to an embodiment of the present invention may mean an artificial intelligence model in which optimization learning for a specific purpose has been performed.

[0129] In other words, a specialized model (SM) may be an artificial intelligence model trained according to specialized training data and methods to achieve a given purpose.

[0130] In an embodiment, such a specialized model (SM) may include a learned predetermined sLLM (including MoELM and / or DMoE models), a general MoE model (NM), and / or an external model (EM). In addition, the specialized model (SM) may include a specialized module model according to an embodiment of the present invention disclosed in [MoE-based model specification method] described below. A detailed description thereof is described in the above [MoE-based model specification method].

[0131] In an embodiment, the sLLM, general MoE model (NM), external model (EM) and / or expert model (SM) as described above may be a secondary model (S: Secondary Model) that can perform specific tasks under the control and management of a master model (P) of an AI agent model (i.e., an orchestrator (OCT) and / or a router (RT) etc.).

[0132]

[0133] [How to provide the MoE model]

[0134] Hereinafter, a method for implementing a model provision service based on an MoE (Mixture of Experts) architecture that provides an MoE model based on a plurality of independently learned expert models (SM) by a computing system (1000) according to one embodiment of the present invention will be described in detail with reference to the attached drawings.

[0135] FIG. 5 illustrates a flowchart for explaining a method for providing an MoE model according to one embodiment of the present invention, and FIG. 6 illustrates a conceptual diagram for explaining a method for providing an MoE model according to one embodiment of the present invention.

[0136] Referring to FIGS. 5 and 6, a method for implementing a MoE architecture-based model provision service that provides a MoE model implemented based on a plurality of pre-learned expert models (SM) by a computing system (1000) according to an embodiment of the present invention may include a step (S101) of acquiring a plurality of expert models (SM), a step (S103) of acquiring a predetermined router (RT), a step (S105) of combining the acquired plurality of expert models (SM) and the router (RT), a step (S107) of constructing an MoE model according to the combination of the plurality of expert models (SM) and the router (RT), and a step (S109) of providing output data based on the constructed MoE model.

[0137] Specifically, a computing system (1000) according to one embodiment of the present invention can obtain a plurality of specialized models (SM). (S101)

[0138] Here, in other words, the specialized model (SM) according to the embodiment may mean an artificial intelligence model that has undergone optimization learning for a specific purpose, and that has been learned according to training data and methods specialized for that purpose.

[0139] In an embodiment, such expert model (SM) may include a learned predetermined sLLM (including MoELM and / or DMoE model), a general MoE model (NM), an external model (EM), and / or an expert module model according to an embodiment of the present invention disclosed in [MoE-based model specification method] described below.

[0140] In detail, in an embodiment, the computing system (1000) can obtain a plurality of expert models (SM) as described above based on user input and / or connection with an external server.

[0141] In addition, a computing system (1000) according to one embodiment of the present invention can acquire a predetermined router (RT). (S103)

[0142] Here, in other words, the router (RT) according to the embodiment may be an artificial intelligence module that performs task assignment and / or traffic adjustment for multiple models within the MoE architecture.

[0143] In detail, in an embodiment, the computing system (1000) may acquire at least one router (RT) based on user input and / or connection with an external server.

[0144] At this time, the router (RT) may include a router (RT) that has been previously learned according to a disclosed predetermined algorithm, a router (RT) that has been additionally learned according to an embodiment of the present invention described below, and / or a router (RT) that has been newly learned in a new manner.

[0145] In addition, a computing system (1000) according to one embodiment of the present invention can combine a plurality of acquired specialized models (SM) and routers (RT). (S105)

[0146] That is, in the embodiment, the computing system (1000) can mutually combine multiple independently learned expert models (SM) and routers (RT).

[0147] In detail, in an embodiment, a computing system (1000) can build a work model (hereinafter, MoELM) that operates like the MoE architecture of FIG. 6 by adding a router (RT) to a plurality of specialized models (SM).

[0148] In other words, the computing system (1000) can configure a third MoE utilizing a modularized specialized model (SM).

[0149] At this time, in the embodiment, the computing system (1000) can detect the input data type for each of the plurality of specialized models (SM).

[0150] And the computing system (1000) can perform different types of combining processes depending on whether the detected input data types are identical.

[0151] In more detail, in an embodiment, the computing system (1000) can combine multiple specialized models (SM) and routers (RT) by utilizing the existing router (RT) as is when the input data types between multiple specialized models (SM) are the same.

[0152] Specifically, in an embodiment, the computing system (1000) can perform combining of multiple specialized models (SM) and routers (RT) by integrating the weight values ​​of existing routers (RT) for each specialized model (SM) through a predetermined operation.

[0153] As an example, the computing system (1000) can perform combining of multiple specialized models (SM) and routers (RT) by using a method of simply adding up the weight values ​​of existing routers (RT) for each specialized model (SM), a method of sampling and adding up, and / or a method of adding up according to a specific formula according to user settings.

[0154] Through this, the computing system (1000) can combine multiple independently trained expert models (SM) and routers (RT) into one without separate additional learning.

[0155] On the other hand, in the embodiment, the computing system (1000) can combine multiple specialized models (SM) and routers (RT) by additionally learning an existing router (RT) when the input data types of multiple specialized models (SM) are different from each other.

[0156] In the following, for the sake of effective explanation, the multiple expert models (SM) are limited to the first expert model (SM) and the second expert model (SM), but are not limited thereto.

[0157] In detail, the computing system (1000) as an embodiment can train the router (RT) to expand the input data types of the first specialized model (SM) and the second specialized model (SM).

[0158] In more detail, the computing system (1000) can integrate the original form of input data of the first specialized model (SM) and the original form of input data of the second specialized model (SM) with each other according to a predetermined operation while preserving them.

[0159] And the computing system (1000) can train the router (RT) to use the input data type according to the integrated input data as the input data type of the first specialized model (SM) and the second specialized model (SM).

[0160] In other words, the computing system (1000) can perform additional learning on the router (RT) to expand the input data types of the first specialized model (SM) and the second specialized model (SM) and perform learning and operations based thereon.

[0161] In another embodiment, the computing system (1000) can train the router (RT) to change the input data types of the first specialized model (SM) and the second specialized model (SM).

[0162] In detail, the computing system (1000) can transform the input data of the first specialized model (SM) into a weight value and transform it into an input data type of the second specialized model (SM).

[0163] And the computing system (1000) can train the router (RT) to learn and operate the second expert model (SM) using the transformed data values.

[0164] In the same manner, the computing system (1000) can train the router (RT) to transform the input data of the second specialized model (SM) into the input data type of the first specialized model (SM) and then use it to learn and operate the first specialized model (SM).

[0165] That is, the computing system (1000) can perform additional learning on the router (RT) to change the input data types of the first specialized model (SM) and the second specialized model (SM) and perform learning and operation accordingly.

[0166] In this way, in the embodiment, the computing system (1000) can build a combined model (i.e., MoELM) that prevents errors and performs smoother learning and operation by combining multiple expert models (SM) and routers (RT) with additional learning that resolves differences in input data types between expert models (SM).

[0167] Meanwhile, according to an embodiment, the computing system (1000) can combine multiple expert models (SM) and routers (RT) by learning an existing router (RT) in a new way.

[0168] In detail, the computing system (1000) as an embodiment can perform router (RT) learning to determine which input data to pass on to which expert model (SM) based on input data and / or available resources.

[0169] At this time, the computing system (1000) may perform router (RT) learning to determine whether to transmit all or only part of the input data.

[0170] Additionally, the computing system (1000) as an embodiment can perform router (RT) learning to determine how many expert models (SM) to activate based on input data and / or available resources.

[0171] Additionally, the computing system (1000) as an embodiment can perform router (RT) learning to judge the quality of output data and accordingly determine which input data to pass on to which expert model (SM).

[0172] Additionally, the computing system (1000) as an embodiment may perform router (RT) learning to determine data processing speed and / or performance and, accordingly, determine which input data to pass on to which expert model (SM).

[0173] In addition, the computing system (1000) as an embodiment can perform router (RT) learning to determine the quality of data processing according to the initial setting and, accordingly, determine which input data to transmit to which expert model (SM).

[0174] Additionally, as an embodiment, the computing system (1000) can perform router (RT) learning to determine which input data to pass to which specialized model (SM) based on target performance and data processing speed according to user input.

[0175] In addition, the computing system (1000) as an embodiment can perform router (RT) learning in which the router (RT) determines the optimal expert model (SM) on its own based on user feedback, reinforcement learning, LLM-based self-feedback, RLHF and / or RLAIF, and transmits input data to the corresponding expert model (SM).

[0176] In addition, referring further to FIG. 4, according to an embodiment, the computing system (1000) may further include an artificial intelligence module (hereinafter, MoE Automatic Setting Module (ASM)) that learns a predetermined existing MoE architecture and automatically sets a new MoE based on the same.

[0177] And the computing system (1000) can set learning and operation for the router (RT) using the above-described MoE automatic setting module (ASM).

[0178] At this time, the MoE Auto Setting Module (ASM) may not perform any additional operations after completing the settings for the router (RT).

[0179] Additionally, according to an embodiment, the computing system (1000) may perform router (RT) learning that implements the functional operation of the MoE automatic setting module (ASM) described above.

[0180] In this way, the computing system (1000) in the embodiment can execute a variety of new router (RT) learning processes so that multiple specialized models (SM) and routers (RT) can be seamlessly combined with each other.

[0181] Thus, the computing system (1000) can more stably guarantee the performance and quality of the combined model (i.e., MoELM).

[0182] In addition, a computing system (1000) according to one embodiment of the present invention can construct an MoE model by combining multiple specialized models (SM) and routers (RT). (S107)

[0183] That is, in the embodiment, the computing system (1000) can build a work model (i.e., MoELM) that operates like the MoE architecture by combining a plurality of specialized models (SM) and routers (RT) as described above.

[0184] At this time, according to the embodiment, MoELM may be included in sLLM according to the embodiment of the present invention.

[0185] In other words, the sLLM according to the embodiment may include the MoELM constructed according to the embodiment of the present invention.

[0186] Additionally, the computing system (1000) according to one embodiment of the present invention can provide output data based on the constructed MoE model. (S109)

[0187]

[0188] *That is, in the embodiment, the computing system (1000) can provide output data (e.g., response data for a specific query and / or a control signal according to a specific command) for a predetermined input data (e.g., text, voice, image, video, and / or sensing data based on a specific sensor, etc.) using the MoELM constructed as above.

[0189] As described above, in the embodiment, the computing system (1000) can operate like an MoE architecture by combining separately learned expert models (SM) into a router (RT) and provide output data accordingly.

[0190] In other words, the computing system (1000) can generate and provide output data according to task processing based on predetermined input data by using a combined model (i.e., MoELM) constructed in the form of MoE by collecting independently learned single models.

[0191] Through this, in the embodiment, the computing system (1000) can implement an MoE mechanism that implements highly efficient data processing without wasting unnecessary resources by utilizing small-sized models optimized for a given task, unlike the existing MoE method that operates while maintaining a large-sized model overall.

[0192] Thus, the computing system (1000) can support various services using an improved MoE model that maintains the advantages of the conventional MoE method (e.g., faster pre-learning speed compared to FFN, faster inference speed compared to models of the same size, and / or improved Instruction Tuning performance, etc.) while minimizing the disadvantages of the conventional MoE method (e.g., the need for a high level of VRAM and / or various challenges in the Fine Tuning process, etc.), and thus can directly improve the performance and quality of the corresponding service.

[0193]

[0194] [MoE-based model specification method]

[0195] Hereinafter, a method for implementing a model provision service based on MoE architecture that implements modularization for a predetermined expert model (SM) within an MoE (Mixture of Experts) model by a computing system (1000) according to one embodiment of the present invention will be described in detail with reference to the attached drawings.

[0196] FIG. 7 illustrates a flowchart for explaining a model specification method based on MoE according to one embodiment of the present invention, and FIG. 8 illustrates a conceptual diagram for explaining a model specification method based on MoE according to one embodiment of the present invention.

[0197] Referring to FIGS. 7 and 8, a method for implementing a model provision service based on MoE architecture that performs modularization of a specialized model (SM) included in an MoE model by a computing system (1000) according to an embodiment of the present invention may include a step (S201) of performing MoE learning based on MoELM, a step (S203) of acquiring specialized model (SM) characteristic information according to MoE learning, a step (S205) of generating a specialized module model based on the acquired specialized model (SM) characteristic information, a step (S207) of acquiring predetermined domain information, a step (S209) of determining a domain-specific specialized model based on the acquired domain information, a step (S211) of constructing a MoE model based on the determined domain-specific specialized model, and a step (S213) of providing output data based on the constructed MoE model.

[0198] Specifically, in many cases, it is difficult to distinguish or identify which domain a general trained expert model (SM) is specialized for.

[0199] This may result in certain limitations in selecting and utilizing the most optimized specialized model (SM) for a specific task.

[0200] To resolve this, in an embodiment of the present invention, the computing system (1000) can perform the following process of specifying the role and / or function of each specialized model (SM) and modularizing it, and effectively selecting and utilizing a customized specialized model (SM) optimized for a specific domain based on the same.

[0201] In detail, a computing system (1000) according to one embodiment of the present invention can perform MoELM-based MoE learning. (S201)

[0202] That is, in the embodiment, the computing system (1000) can perform MoE learning based on the MoELM constructed according to the above-described [MoE model provision method].

[0203] At this time, as learning is performed, the computing system (1000) can implement learning for each of the multiple expert models (SM) included in MoELM.

[0204] In other words, as learning as above is performed, multiple expert models (SMs) within MoELM can be trained individually.

[0205] Here, in other words, the specialized model (SM) according to the embodiment may mean an artificial intelligence model that has undergone optimization learning for a specific purpose, and that has been trained according to training data and methods specialized for that purpose.

[0206] In an embodiment, such expert models (SM) may include a learned predetermined sLLM (including MoELM and / or DMoE models), a general MoE model (NM), an external model (EM), and / or an expert module model (MM) according to an embodiment of the present invention disclosed below.

[0207] In addition, a computing system (1000) according to one embodiment of the present invention can obtain specialized model feature information (SMFI) according to MoE learning. (S203)

[0208] Here, the specialized model feature information (SMFI) according to the embodiment may mean information that specifies the role and / or function of a given specialized model (SM).

[0209] In detail, referring further to FIG. 4, in an embodiment, the computing system (1000) may further include a model specialization module (MSM) according to an embodiment of the present invention.

[0210] And the computing system (1000) can obtain the specialized model feature information (SMFI) described above through linkage with the model specification module (MSM).

[0211] Here, the model specialization module (MSM) according to an embodiment of the present invention may be an artificial intelligence module that generates and outputs specialized model feature information (SMFI) corresponding to a given specialized model (SM) based on MoE learning.

[0212] Specifically, in the embodiment, the model specification module (MSM) can monitor and track the task assignment status of the router (RT) for each specialized model (SM) when the above-described MoE learning is performed.

[0213] That is, in the embodiment, the model specification module (MSM) can determine which task the router (RT) distributes and assigns to which specialized model (SM) as the MoELM learns and operates.

[0214] Depending on the embodiment, the model specification module (MSM) may also manage matching by creating a tag that specifies each tracked task assignment status.

[0215] Through this, in the embodiment, the model specialization module (MSM) can determine the expertise of each of the multiple expert models (SM).

[0216] Additionally, in the embodiment, the model specification module (MSM) can generate specialized model feature information (SMFI) corresponding to each specialized model (SM) based on the expertise of each determined specialized model (SM).

[0217] FIG. 9 illustrates an example of specialized model feature information (SMFI) according to one embodiment of the present invention.

[0218] At this time, referring to FIG. 9, the model specification module (MSM) as an embodiment can generate the specialized model feature information (SMFI) described above in at least one of the following formats.

[0219] [Format 1] Specialized model feature information (SMFI) in the form of selecting one of the preset specialized model (SM) roles and / or function-specific categories (e.g., question-answering or device control) based on user input.

[0220]

[0221] *[Type 2] Specialized model feature information (SMFI) in the form of natural language that specifies the role and / or function of the specialized model (SM).

[0222] [Third Format] Specialized Model Characteristic Information (SMFI) in the form of specifying the role and / or function of the specialized model (SM) in at least one of the first and second formats, and further defining the input data and output data of the specialized model (SM).

[0223] Continuing, in the embodiment, the model specification module (MSM) can provide the specialized model feature information (SMFI) generated as above as output data to the computing system (1000).

[0224] Thus, in the embodiment, the computing system (1000) can obtain characteristic information for each specialized model (SM) through linkage with a model specification module (MSM).

[0225] In addition, a computing system (1000) according to one embodiment of the present invention can generate a specialized module model (MM) based on acquired specialized model feature information (SMFI). (S205)

[0226] Here, the specialized module model (MM) according to the embodiment of the present invention may mean an independently separated specialized model (SM) with matching specialized model feature information (SMFI).

[0227] In detail, in an embodiment, the computing system (1000) can match the acquired specialized model feature information (SMFI) as described above to a corresponding specialized model (SM).

[0228] Additionally, in the embodiment, the computing system (1000) can independently separate and database each expert model (SM) with matched expert model feature information (SMFI).

[0229] That is, in the embodiment, the computing system (1000) can perform modularization by matching each specialized model (SM) with corresponding specialized model characteristic information (SMFI) and storing and managing them separately.

[0230] Thus, the computing system (1000) can generate a specialized module model (MM), which is an independently separated specialized model (SM), while matching the specialized model feature information (SMFI).

[0231] In this way, in the embodiment, the computing system (1000) can identify the characteristics of each specialized model (SM) within a given MoE model (MoELM in the embodiment) and, by reflecting these characteristics, modularize each specialized model (SM) into a small size that can be reused and shared.

[0232] Accordingly, the computing system (1000) can quickly and efficiently select and screen a specialized model (SM) that implements a data processing process optimized for a specific domain with higher accuracy, and can easily support flexible expansion or reduction of the MoE model based thereon.

[0233] In addition, the computing system (1000) according to one embodiment of the present invention can obtain predetermined domain information. (S207)

[0234] Here, domain information according to an embodiment may be information defining a domain that specifies data, rules, terms, problem definitions, and / or processes that a given AI system uses to perform a given task.

[0235] In detail, in an embodiment, the computing system (1000) may obtain certain input data (e.g., text, voice, image, video, and / or sensing data based on a specific sensor, etc.).

[0236] Additionally, in the embodiment, the computing system (1000) can determine a domain corresponding to the acquired input data.

[0237] At this time, the method by which the computing system (1000) determines the domain for input data in the embodiment can be performed based on various disclosed algorithms capable of performing the same, and the embodiment of the present invention does not limit or restrict the algorithm itself.

[0238] Thus, in the embodiment, the computing system (1000) can obtain domain information about a task to be processed.

[0239] Additionally, a computing system (1000) according to one embodiment of the present invention can determine a domain-specific specialized model based on acquired domain information. (S209)

[0240] Here, the domain-specific specialized model according to the embodiment may mean a specialized model (SM) that executes data processing (e.g., deep learning, etc.) operations optimized for a given domain.

[0241] In detail, referring further to FIG. 8, in an embodiment, the computing system (1000) can determine at least one domain-specific specialized model based on the domain information and specialized model feature information (SMFI) acquired as described above.

[0242] In more detail, in an embodiment, the computing system (1000) can detect at least one specialized model feature information (SMFI) having a characteristic corresponding to the acquired domain information.

[0243] For example, when the computing system (1000) confirms the 'characteristics of a task that outputs response data for a given query data' based on the first domain information, it can detect at least one specialized model characteristic information (SMFI) that is specified as a 'role and / or function specialized in query response' among a plurality of specialized model characteristic information (SMFI) stored in a database.

[0244] At this time, according to an embodiment, the computing system (1000) can detect at least one specialized model feature information (SMFI) corresponding to domain information based on a plurality of tags generated by a model specification module (MSM) for each task allocation state of a router (RT) for a plurality of specialized models (SM) during learning based on the above-described MoE architecture.

[0245] That is, according to an embodiment, the computing system (1000) can detect at least one specialized model feature information (SMFI) corresponding to the domain information by comparing the plurality of tags and domain information generated as described above.

[0246] At this time, according to an embodiment, the computing system (1000) can filter the tags to be compared according to the creation time of each tag.

[0247] Specifically, the computing system (1000) can set at least one tag generated at a specific task assignment time as a comparison target tag according to user input and / or a preset self-process.

[0248] For example, the computing system (1000) may set at least one tag generated for a task assignment status performed after a preset time during the entire learning time as a comparison target tag, taking into account that task assignment accuracy improves as the learning rate increases.

[0249] Thus, the computing system (1000) can detect at least one specialized model feature information (SMFI) corresponding to the domain information through a comparison between at least one filtered tag that ensures higher accuracy and the domain information.

[0250] Additionally, in the embodiment, the computing system (1000) can extract a specialized model (SM) (i.e., a specialized module model (MM)) that matches each of at least one detected specialized model feature information (SMFI).

[0251] And in the embodiment, the computing system (1000) can determine at least one extracted specialized module model (MM) as a domain-specific specialized model.

[0252] In addition, a computing system (1000) according to one embodiment of the present invention can build an MoE model based on a determined domain-specific expert model. (S211)

[0253] Referring further to FIG. 8, that is, in the embodiment, the computing system (1000) can build a work model (hereinafter, DMoE model) that operates like the MoE architecture based on at least one domain-specific specialized model determined as above.

[0254] In other words, the computing system (1000) can build an MoE model (i.e., a DMoE model) that implements data processing optimized for a specific domain by using at least some models (i.e., domain-specific specialized models) among a plurality of specialized models (SM) modularized into small sizes.

[0255] In detail, in an embodiment, the computing system (1000) can build the DMoE model described above by combining at least one domain-specific specialized model and a predetermined router (RT).

[0256] At this time, in the embodiment, the specific method of building a DMoE model by combining a domain-specific specialized model and a router (RT) by the computing system (1000) may be applied to the description of the method of building a MoELM by combining a plurality of specialized models (SM) and routers (RT) disclosed in the above-described [MoE model provision method].

[0257] Thus, in the embodiment, the computing system (1000) can build a DMoE model including a domain-specific specialized model and a router (RT).

[0258] At this time, according to the embodiment, the DMoE model may be included in the sLLM according to the embodiment of the present invention.

[0259] In other words, the sLLM according to the embodiment may include a DMoE model constructed according to the embodiment of the present invention.

[0260] Additionally, the computing system (1000) according to one embodiment of the present invention can provide output data based on the constructed MoE model. (S213)

[0261] That is, in the embodiment, the computing system (1000) can provide output data (e.g., response data for a specific query and / or a control signal according to a specific command) for a predetermined input data (e.g., text, voice, image, video, and / or sensing data based on a specific sensor, etc.) using the DMoE model constructed as described above.

[0262] As described above, in the embodiment, the computing system (1000) can specify the role and / or function of each specialized model (SM) and separate and modularize them to a level where they can be reused and shared, and can utilize this to quickly and flexibly build a customized MoE model (i.e., a DMoE model) optimized for a specific domain, and can provide predetermined output data according to efficient task processing using the built model.

[0263] In other words, in the embodiment, the computing system (1000) can implement and provide an MoE model with improved data processing (and / or operation) speed and inference performance, and can support various services through the same, thereby effectively promoting improvement in performance and quality.

[0264]

[0265] [Method for Providing AI Agents Based on MoE Application LLM]

[0266] Hereinafter, a method for implementing a model provision service based on MoE architecture that determines an application model optimized for a domain according to an external environment based on an LLM (Large Language Model) that applies MoE (Mixture of Experts) and provides an on-device specialized AI agent (Artificial Intelligence Agent) that performs output based on the determined application model will be described in detail with reference to the attached drawings, according to an embodiment of the present invention, by a computing system (1000).

[0267] FIG. 10 illustrates a flowchart for explaining a method for providing an AI agent based on an MoE application LLM according to one embodiment of the present invention, and FIG. 11 illustrates a conceptual diagram for explaining a method for providing an AI agent based on an MoE application LLM according to one embodiment of the present invention.

[0268] Referring to FIGS. 10 and 11, a method for implementing a model provision service based on MoE architecture, which provides an on-device specialized AI agent specialized model (AIAM) that determines an application model optimized for a domain according to an external environment based on an LLM that applies MoE according to an embodiment of the present invention, and performs an output based on the determined application model, may include a step (S301) of executing an on-device AI agent service, a step (S303) of obtaining predetermined input data, a step (S305) of determining a domain according to the obtained input data, a step (S307) of determining an application model according to the determined domain, and a step (S309) of providing output data based on the determined application model.

[0269] Specifically, a computing system (1000) according to one embodiment of the present invention can execute an on-device AI agent service. (S301)

[0270] Here, for reference, on-device AI refers to technology that performs AI-based data processing directly on the user's device, rather than on the cloud and / or external servers. This can offer benefits such as privacy protection, real-time processing, and reduced dependence on internet connectivity, as all processing is completed on the device without sending data externally.

[0271] Therefore, in this regard, on-device AI agent service can refer to various services implemented using on-device AI.

[0272] For example, on-device AI agent services may include voice assistant services for smartphones (e.g., Google Assistant, Apple Siri, Samsung Bixby, etc.), smart camera services (e.g., Google Pixel's HDR+, Apple's Deep Fusion, etc.), fitness tracker and smartwatch services (e.g., Apple Watch, Fitbit, etc.), autonomous driving services for automobiles (e.g., Tesla's Autopilot, etc.), and / or home security services (e.g., Nest Secure, Ring, etc.).

[0273] In an embodiment, the computing system (1000) may execute a predetermined on-device AI agent service based on an AI agent specialization model (AIAM) according to an embodiment of the present invention and / or linkage with a predetermined application, etc.

[0274] In addition, the computing system (1000) according to one embodiment of the present invention can obtain predetermined input data. (S303)

[0275] In detail, in the embodiment, the computing system (1000) can obtain at least one input data (e.g., predetermined text, voice, image, video and / or sensing data, etc.) based on the user input and / or linkage with an external device (e.g., predetermined sensor, etc.) based on the on-device AI agent service executed as described above.

[0276] In an embodiment, the input data obtained as above may include certain data that can specify the target task of data processing.

[0277] In addition, the computing system (1000) according to one embodiment of the present invention can determine a domain based on acquired input data. (S305)

[0278] Here, again, the domain according to the embodiment may mean data, rules, terminology, problem definitions, and / or processes that a given AI system uses to perform a specific task.

[0279] In detail, in an embodiment, the computing system (1000) can determine a domain corresponding to acquired input data.

[0280] At this time, the method by which the computing system (1000) determines the domain for input data in the embodiment can be performed based on various disclosed algorithms capable of performing the same, and the embodiment of the present invention does not limit or restrict the algorithm itself.

[0281] Thus, in the embodiment, the computing system (1000) can obtain domain information corresponding to the task to be processed.

[0282] In addition, the computing system (1000) according to one embodiment of the present invention can determine an application model according to the determined domain. (S307)

[0283] Here, the application model according to the embodiment may mean a model that performs a predetermined task processing according to given input data.

[0284] In an embodiment, such an application model may be at least one of the secondary models (S) described above.

[0285] At this time, in other words, the secondary model (S) according to the embodiment may mean a model that can perform a specific task according to the control and management of the master model (P) (i.e., the orchestrator (OCT) and / or router (RT), etc.) responsible for controlling and managing the operation of a given AI system.

[0286] In an embodiment, such secondary model (S) may include at least one of an sLLM (including a MoELM and / or DMoE model), a general MoE model (NM), an external model (EM) and / or a specialized model (SM) (including a specialized module model (MM)).

[0287] In detail, in an embodiment, the computing system (1000) can determine at least one application model based on the domain information acquired as described above.

[0288] In more detail, the computing system (1000) according to the embodiment of the present invention can detect at least one model (i.e., domain-specific model) that executes data processing (e.g., deep learning, etc.) operation optimized for given domain information among the secondary models (S) described above in conjunction with the master model (P) according to the embodiment of the present invention (i.e., orchestrator (OCT) and / or router (RT), etc.).

[0289] At this time, in the embodiment of the present invention, the specific method by which the computing system (1000) detects a domain-specific model in conjunction with the master model (P) is omitted by applying the description of the router (RT) and orchestrator (OCT) disclosed in the above-described [AI Agent Specialization Model (AIAM)].

[0290] Additionally, in the embodiment, the computing system (1000) can determine at least one detected domain-specific model as an applicable model.

[0291] Additionally, the computing system (1000) according to one embodiment of the present invention can provide output data based on the determined application model. (S309)

[0292] That is, in the embodiment, the computing system (1000) can generate and provide output data (e.g., response data for a specific query and / or control signals according to a specific command) for a given input data (e.g., text, voice, image, video, and / or sensing data based on a specific sensor, etc.) based on at least one application model determined through the AI ​​agent specialization model (AIAM) as described above.

[0293] In other words, the computing system (1000) can perform a predetermined request task based on given input data using the application model determined as above, and provide output data according to the performed data processing.

[0294] At this time, in the embodiment, the computing system (1000) can provide the output data based on the on-device AI agent service described above.

[0295] As described above, in the embodiment, the computing system (1000) can effectively determine a model optimized for data processing according to a given domain even in an on-device environment based on an AI agent specialized model (AIAM) including models implemented by applying the MoE architecture according to the embodiment of the present invention (for example, MoELM, DMoE model and / or specialized module model (MM) etc.), and provide an output according to efficient data processing through the determined model.

[0296] That is, the computing system (1000) can implement and provide an artificial intelligence model (i.e., an AI agent specialization model (AIAM)) that better understands, better executes, and better responds to a given task in any environment.

[0297] Thus, in the embodiment, the computing system (1000) can directly and significantly improve the quality and performance of various AI agent-based services (e.g., a voice assistant service for a smartphone, a smart camera service, a fitness tracker and smart watch service, an autonomous driving car service, and / or a home security service, etc.).

[0298]

[0299] The embodiments of the present invention described above may be implemented in the form of program commands that can be executed through various computer components and recorded on a computer-readable recording medium. The computer-readable recording medium may include program commands, data files, data structures, etc., either singly or in combination. The program commands recorded on the computer-readable recording medium may be specially designed and configured for the present invention or may be known and available to those skilled in the art of computer software. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tapes, optical recording media such as CD-ROMs and DVDs, magneto-optical media such as floptical disks, and hardware devices specifically configured to store and execute program commands, such as ROMs, RAMs, and flash memories. Examples of program commands include not only machine language codes generated by a compiler, but also high-level language codes that can be executed by a computer using an interpreter, etc. Hardware devices may be changed into one or more software modules to perform processing according to the present invention, and vice versa.

[0300] The specific implementations described in the present invention are exemplary embodiments and do not limit the scope of the present invention in any way. For the sake of brevity, descriptions of conventional electronic components, control systems, software, and other functional aspects of the systems may be omitted. In addition, the lines connecting or connecting members between components illustrated in the drawings are merely representative of functional connections and / or physical or circuit connections, and may be replaced or represented as various additional functional connections, physical connections, or circuit connections in an actual device. In addition, unless specifically mentioned as “essential,” “important,” etc., a component may not be absolutely necessary for the application of the present invention.

[0301] Although the detailed description of the present invention has been described with reference to preferred embodiments of the present invention, it will be understood by those skilled in the art or having ordinary knowledge in the art that various modifications and changes can be made to the present invention without departing from the spirit and technical scope of the present invention as set forth in the claims below. Accordingly, the technical scope of the present invention should not be limited to the contents described in the detailed description of the specification, but should be defined by the claims.

[0302]

[0303] The present invention relates to a method and system for providing an artificial intelligence model including a plurality of models, and is applicable to the artificial intelligence industry, and thus has industrial applicability.

Claims

1. A method for providing an artificial intelligence model including multiple models by a computing system including a memory and a processor, A step of acquiring multiple specialized models, which are independently trained artificial intelligence models; Step of acquiring a router (Gating Network), which is an artificial intelligence module that controls model tasks; A step of combining the above-mentioned acquired router and multiple specialized models; A step of building MoELM, which is a model based on MoE (Mixture of Experts) architecture according to the above combination; and A step of providing output data for a predetermined input data based on the above-constructed MoELM. A method for providing an artificial intelligence model including multiple models.

2. In paragraph 1, The above router, Performing task assignment for each of the plurality of expert models based on the above input data A method for providing an artificial intelligence model including multiple models.

3. In paragraph 2, The step of combining the above router and multiple specialized models is: A step of detecting an input data type for each of the above multiple specialized models, Including a step of performing a different type of combination process depending on whether the detected input data types are identical or not. A method for providing an artificial intelligence model including multiple models.

4. In paragraph 3, The steps for performing the above combining process are: If the input data types between the plurality of specialized models are identical, a step of performing a combination between the router and the plurality of specialized models using the acquired router as is is included. A method for providing an artificial intelligence model including multiple models.

5. In paragraph 4, The step of performing a combination between the router and multiple specialized models using the above-mentioned acquired router as is, A step of integrating the weight values ​​of the router for each of the plurality of specialized models according to a predetermined operation. A method for providing an artificial intelligence model including multiple models.

6. In paragraph 3, The steps for performing the above combining process are: If the input data types between the plurality of specialized models are different from each other, a step of performing combination between the router and the plurality of specialized models by additionally learning the acquired router is included. A method for providing an artificial intelligence model including multiple models.

7. In paragraph 6, The step of performing additional learning on the acquired router and combining the router and multiple expert models is as follows: A step of performing additional learning to expand the input data type of each of the plurality of expert models is included. A method for providing an artificial intelligence model including multiple models.

8. In paragraph 7, The step of performing additional learning to expand the input data type of each of the above multiple expert models is as follows: A step of mutually integrating the input data for each of the plurality of specialized models according to a predetermined operation while preserving the original form of the input data for each of the plurality of specialized models, A step of performing additional learning to set the input data type according to the integrated input data to the input data type of each of the plurality of expert models. A method for providing an artificial intelligence model including multiple models.

9. In paragraph 6, The step of performing additional learning on the acquired router and combining the router and multiple expert models is as follows: A step of performing additional learning to change the input data type of each of the plurality of expert models is included. A method for providing an artificial intelligence model including multiple models.

10. In paragraph 9, The step of performing additional learning to change the input data type of each of the above multiple expert models is as follows: A step of converting the input data of a predetermined first specialized model into a weight value and transforming it into the input data type of a predetermined second specialized model, A step of performing additional learning to learn and operate the second expert model based on the transformed data values. A method for providing an artificial intelligence model including multiple models.

11. In paragraph 2, The step of combining the above router and multiple specialized models is: A step of performing additional learning on the acquired router to learn a predetermined existing MoE architecture and automatically set up a new MoE based on the learned existing MoE architecture; A step of performing a combination between the router and a plurality of expert models based on the additionally learned router. A method for providing an artificial intelligence model including multiple models.

12. At least one memory; and At least one processor for providing an artificial intelligence model including a plurality of models by reading out at least one application stored in the memory; The instructions of the above processor are: A step of acquiring multiple specialized models, which are independently trained artificial intelligence models, and The step of acquiring a router (Gating Network), which is an artificial intelligence module that controls the model task, A step of combining the above-mentioned acquired router and multiple specialized models, A step for building MoELM, a model based on MoE (Mixture of Experts) architecture according to the above combination, Includes a command for performing a step of providing output data for a predetermined input data based on the above-constructed MoELM. An artificial intelligence model providing system including multiple models.

Images