Method and system for configuring distributed federated learning topology for on-device ai network system

Abstract

The present invention relates to an unmanned spaceport and space-Earth flight vehicle operating system and an unmanned space-Earth flight vehicle operating method, the system comprising: an unmanned spaceport provided and operated on the ground; a takeoff and landing platform provided in the unmanned spaceport; an aerial support unit capable of flying from the takeoff and landing platform to a predetermined altitude point; a space-Earth flight vehicle taking off or landing on the aerial support unit; and an energy supply means for supplying power or fuel between the unmanned spaceport and the aerial support unit. Therefore, the unmanned spaceport and space-Earth flight vehicle operating system is intended to enable mass production and global standardization through a two-stage operating structure separated into the ground and the air, a fuel and electricity supply method in the air, a function for maintaining a platform airborne for an extended period by means of a drone-type aerial support unit, modular flight vehicles, and an airport structure.

Description

Method and System for Configuring a Distributed Federated Learning Topology for On-Device AI Network Systems

[0001] Cross-reference regarding related applications

[0002] This application claims the benefit of priority to Korean Patent Application No. 10-2025-0013224 filed on February 3, 2025, the full text of said patent application incorporated by reference into this specification.

[0003] The embodiments disclosed in this document relate to a distributed federated learning topology configuration method and system for an on-device AI network system.

[0004] With the recent explosive growth of smartphones and the Internet of Things, vast amounts of data are being generated across various fields, and as the size of this data increases, centralized, cloud / server-centric AI networks face the problem of difficulty in providing rapid and seamless services due to network latency and increased server load.

[0005] In addition, concerns regarding privacy, security, and data governance are being raised in existing cloud / server-centric, centralized AI networks because sensitive data is collected in a single repository.

[0006] To address these issues, on-device AI technology, which involves embedding and processing AI models directly into the device itself, is gaining attention.

[0007] Currently, dedicated lightweight AI model architectures and model compression technologies are being developed in stand-alone on-device AI network environments, and various AI models utilizing these technologies are being introduced; however, existing on-device AI technologies face the problem of performance limitations due to constraints on computing resources and small-scale data.

[0008] According to one embodiment disclosed in this document, a method and system for configuring a distributed federated learning topology for an on-device AI network system are provided, which can improve the performance of a learning model of an on-device AI network system by configuring an optimal distributed federated learning topology for an on-device AI network system.

[0009] According to one embodiment disclosed in this document, a distributed federated learning topology configuration method and system for an on-device AI network system are provided, which can reduce communication traffic and maintain optimal performance by mutually sharing models learned by a plurality of on-device AI nodes constituting the on-device AI network system and evaluating the shared learned models round by round to remove on-device AI nodes with low performance.

[0010] The technical problems of the embodiments disclosed in this document are not limited to those mentioned above, and other unmentioned technical problems will be clearly understood by those skilled in the art from the description below.

[0011] A method for configuring a decentralized federated learning topology in an on-device AI network system that performs decentralized federated learning in rounds, comprising a plurality of on-device AI nodes interconnected according to one embodiment, may comprise: a first step in which each of the plurality of on-device AI nodes learns a local model using local data collected in one round; a second step in which the learned local model is transmitted to at least one connected neighbor node to mutually share the local model; a third step in which the learned local model and the shared local model are evaluated; a fourth step in which a straggler node is identified using the result of the evaluation; a fifth step in which an optimal topology is determined by removing the identified straggler node; and a sixth step in which the mutual connections are reconfigured according to the determined optimal topology.

[0012] At this time, the third step can input the accuracy derived from the waiting time and loss-based performance for each local model into the peer weighting algorithm to derive the output peer weighting matrix as an evaluation result.

[0013] At this time, the peer weighting algorithm may be either Particle Swarm Optimization (PSO) or Hidden Markov Model (HMM).

[0014] In addition, the above fourth step can identify straggler nodes using a Graph Neural Network (GNN).

[0015] At this time, the fourth step inputs the peer weight matrix derived according to the result of performing the evaluation and the adjacency matrix for the topology in which the plurality of on-device AI nodes are interconnected into a graph neural network, obtains a topology weight matrix for the plurality of on-device AI nodes from the graph neural network, and can identify nodes in the obtained topology weight matrix that have a waiting time greater than a certain standard or a loss-based performance less than a certain standard as straggler nodes.

[0016] Additionally, after the above 6th step, until the designated learning round is completed, each of the plurality of on-device AI nodes that are reconfigured and interconnected according to the determined optimal topology may repeat the above 1st through 6th steps.

[0017] Additionally, the present invention may provide a computer-readable recording medium that records a program for executing the method described above.

[0018] According to one embodiment, an on-device AI network system comprising a plurality of on-device AI nodes interconnected according to one topology and performing decentralized federated learning in rounds, wherein the on-device AI nodes may be configured to include: a shared memory that stores a local model learned using local data collected in one round and a shared local model transmitted from at least one connected neighbor node; a model evaluation unit that evaluates the learned local model and the shared local model; and a topology management unit that identifies a straggler node using the result of the evaluation performed by the model evaluation unit and reconstructs the mutual connections according to an optimal topology in which the identified straggler node is removed.

[0019] The above model evaluation unit can input the accuracy derived from the waiting time and loss-based performance for each local model into the peer weighting algorithm and derive the output peer weight matrix as the evaluation result.

[0020] In addition, the topology management unit inputs the peer weight matrix derived according to the result of performing the evaluation and the adjacency matrix for the topology in which the plurality of on-device AI nodes are connected into a graph neural network, obtains a topology weight matrix for each on-device AI node from the graph neural network, and can identify nodes as straggler nodes in the obtained topology weight matrix where the waiting time is above a certain standard or the loss-based performance is below a certain standard.

[0021] According to a distributed federated learning topology configuration method and system for an on-device AI network system according to one embodiment, by removing low-performance on-device AI nodes on a round-by-round basis, latency caused by the overhead of data exchange and communication between on-device AI nodes is reduced, and rapid learning convergence and overall network performance improvement are enabled.

[0022] In addition, according to the present invention, by configuring a distributed federated learning topology optimized for an on-device AI network system, gradient bias and data heterogeneity of the learning model can be overcome, and the stability and effectiveness of the learning model can be improved, and the reliability and robustness to respond to on-device AI node failures can be corrected.

[0023] Through this, optimal distributed learning models can be generated through continuous collaboration of on-device AI nodes without cloud / server intervention, and it becomes possible to develop intelligent autonomous services such as accurate situational awareness, prediction, optimal control, and response on on-device AI nodes.

[0024] FIGS. 1 and FIGS. 2 are illustrative diagrams for explaining an on-device AI network system according to an embodiment of the present invention.

[0025] FIG. 3 is an illustrative diagram for explaining distributed federated learning in an on-device AI network system according to an embodiment of the present invention.

[0026] FIGS. 4 to 15 are drawings for explaining the main configuration of an on-device AI node and a distributed federated learning topology configuration method according to an embodiment of the present invention.

[0027] Hereinafter, various embodiments disclosed in this document will be described in detail with reference to the attached drawings. In this document, the same reference numerals are used for identical components in the drawings, and redundant descriptions of identical components are omitted.

[0028] With respect to the various embodiments disclosed in this document, specific structural or functional descriptions are provided merely for the purpose of explaining the embodiments, and the various embodiments disclosed in this document may be implemented in various forms and should not be interpreted as being limited to the embodiments described in this document.

[0029] Expressions such as "first," "second," "first," or "second" used in various embodiments may modify various components regardless of order and / or importance and do not limit said components. For example, without departing from the scope of the embodiments disclosed herein, the first component may be named the second component, and similarly, the second component may be renamed the first component.

[0030] The terms used in this document are used merely to describe specific embodiments and are not intended to limit the scope of other embodiments. Singular expressions may include plural expressions unless the context clearly indicates otherwise.

[0031] All terms used herein, including technical or scientific terms, may have the same meaning as generally understood by those skilled in the art of the embodiments disclosed herein. Terms defined in commonly used dictionaries may be interpreted as having the same or similar meaning as they have in the context of the relevant technology, and are not to be interpreted in an ideal or overly formal sense unless explicitly defined in this document. In some cases, even terms defined in this document shall not be interpreted to exclude the embodiments disclosed herein.

[0032] Hereinafter, a distributed federated learning topology configuration method and system for an on-device AI network system according to an embodiment of the present invention will be described with reference to the drawings.

[0033] FIGS. 1 and 2 are illustrative diagrams for explaining an on-device AI network system according to an embodiment of the present invention, and FIG. 3 is an illustrative diagram for explaining distributed federated learning in an on-device AI network system according to an embodiment of the present invention.

[0034] First, referring to FIGS. 1 and 2, an on-device AI network system (500) according to an embodiment of the present invention may be configured to include a plurality of on-device AI nodes (100_1, 100_2, … 100_N) interconnected by any one topology. The topology may include various forms such as a ring or a mesh, as shown in FIG. 2.

[0035] In the following description, multiple on-device AI nodes (100_1, 100_2, … 100_N) interconnected in any one topology may be referred to by reference numeral 100 for convenience of explanation, and when describing specific on-device AI nodes separately, they may be referred to by reference numerals 100_1, 100_2, etc.

[0036] Additionally, the on-device AI node (100) of the present invention is a node that performs distributed federated learning that is performed repeatedly in rounds, and may be referred to as a DFL (Decentralized Federated Learning) node or a node in the drawings, but the aforementioned reference numerals, names, etc. are merely for convenience of explanation and do not limit the scope of the present invention.

[0037] The on-device AI node (100) of the present invention may be a device that processes user input, such as a personal computer, terminal, portable telephone, smartphone, handheld device, or wearable device, and may be configured to include a device such as an IoT (Internet of Things) node that can be linked with the device that processes user input. In addition, the on-device AI node (100) of the present invention may be equipped with a device for local training of an artificial neural network model (e.g., a Raspberry Pi device).

[0038] The on-device AI network system (500) composed of the on-device AI node (100) of the present invention is an autonomous on-device AI network system and means a network capable of autonomously collecting data and learning using an AI model (learning model) installed on the device itself, even without going through a cloud / server or having an internet and network connection.

[0039] In addition, the on-device AI network system (500) of the present invention may be provided in a home, a building, or other hospital, and may perform processes such as collecting local data generated in the home, a building, or other hospital, such as data on energy usage, and predicting energy consumption using the collected data, and for such learning and prediction (inference), each on-device AI node (100) is configured to include a deep learning model for learning and inference, and the deep learning model stored in each on-device AI node (100) may be referred to as a local model.

[0040] In particular, the on-device AI node (100) of the present invention may have at least one connected neighbor node as shown in FIG. 3. For example, among a plurality of on-device AI nodes interconnected by any one topology, on-device AI node 2 (100_2), on-device AI node 3 (100_3), and on-device AI node 4 (100_4) are neighbor nodes of on-device AI node 1 (100_1), and the neighbor nodes of on-device AI node 3 (100_3) become the connected on-device AI node 1 (100_1), on-device AI node 2 (100_2), and on-device AI node 4 (100_4).

[0041] Each on-device AI node (100) trains a local model using local data collected in rounds and performs federated learning by collaborating without a central server according to decentralized federated learning, by transmitting the trained local model to neighboring nodes for sharing.

[0042] To perform such distributed federated learning more efficiently, the connection patterns and topology configurations between multiple on-device AI nodes (100) must be optimized. For example, a topology such as a mesh provides fast convergence and high fault tolerance, but it may consume a lot of bandwidth and have poor scalability. A sparse topology such as a ring or a hierarchical structure provides scalability and reduced bandwidth usage, but it has low fault tolerance and may perform poorly on non-IID (Independent and Identically Distributed) data.

[0043] To solve these problems, an on-device AI node (100) according to an embodiment of the present invention evaluates a local model learned in rounds and a local model shared from neighbor nodes, and if a straggler node is identified as a result of the evaluation, reconfigures the connection according to an optimal topology with the identified straggler node removed, thereby increasing the performance and efficiency of the network.

[0044] The above-described process can be performed by each of the plurality of on-device AI nodes (100), and can be performed by any one of the on-device AI nodes, such as a digital twin platform node as shown in FIG. 3, or by any leader node (100_5).

[0045] A leader node (100_5) according to an embodiment of the present invention can perform model management, device management, topology management, coordinator functions, etc., and can store and manage shared memory.

[0046] The main configuration of an on-device AI node and a method for configuring a distributed federated learning topology according to an embodiment of the present invention will be explained in more detail with reference to FIGS. 4 to 15.

[0047] First, referring to FIG. 4, an on-device AI node (100) according to an embodiment of the present invention may be configured to include a shared memory (10), a model evaluation unit (20), and a topology management unit (30). Depending on the implementation method, all on-device AI nodes (100) constituting the on-device AI network system (500) of the present invention may include a shared memory (10) and a model evaluation unit (20), and in the case of a node operating as a leader node, such as a digital twin platform node, may be configured to further include a topology management unit (30). However, it is not limited thereto, and all on-device AI nodes (100) may be configured to include a shared memory (10), a model evaluation unit (20), and a topology management unit (30).

[0048] The shared memory (10) can store a local model learned using local data collected in one round. In this case, the local model may be an artificial neural network model learned through supervised learning, unsupervised learning, semi-supervised learning, or reinforcement learning. Additionally, the shared memory (10) of the present invention can store a shared local model transmitted from at least one connected neighbor node if such a model exists.

[0049] The model evaluation unit (20) performs the role of evaluating the learned local model and the shared local model. In particular, the model evaluation unit (20) according to an embodiment of the present invention can evaluate local models according to a peer weighting algorithm that considers waiting time and local model performance.

[0050] That is, as illustrated in FIG. 5, the on-device AI node (100) accesses its own local dataset for learning essential for distributed learning. It learns a local model using this own local dataset (1), and each on-device AI node (100) can independently learn a local model, which is a deep learning model (2).

[0051] Afterwards, each on-device AI node (100) can receive and share models (M2, M3, M4, …) from neighboring nodes (3), and can evaluate performance based on loss for each local model using validating data.

[0052] Additionally, each on-device AI node (100) calculates the waiting time for a shared local model transmitted from a neighboring node. Here, the waiting time refers to the time spent waiting to receive a local model from a neighboring node after training a local model, and if the time taken to generate the local training model of each node is long, the waiting time may also increase.

[0053] Subsequently, an accuracy matrix derived from the latency and performance of local models is collected (4), the collected latency and accuracy matrices are input into a peer weighting algorithm, and a peer weighting matrix, which is the evaluation result for each local model, is derived through the peer weighting algorithm (5). The peer weighting matrix can be applied to determine the extent to which local models transmitted from each neighbor node contributed to distributed federated learning.

[0054] In addition, the peer weighting algorithm according to an embodiment of the present invention may use either Particle Swarm Optimization (hereinafter referred to as PSO) or Hidden Markov Model (hereinafter referred to as HMM).

[0055] First, the peer weighting algorithm based on the PSO algorithm will be explained with reference to Figures 6 and 7.

[0056] First, in Fig. 6, the accuracy (P) and waiting time (W) derived through the performance as described above are entered as inputs.

[0057] Then, initialize the weight function defined as in the following mathematical formula.

[0058] [Mathematical Formula 1]

[0059]

[0060] Here, α and β are parameters for the particles constituting the population, and each parameter is also initialized.

[0061] In addition, the fitness function (f) for deriving the optimized solution is as follows.

[0062] [Mathematical Formula 2]

[0063]

[0064] The fitness function (f) is intended to quantify how well a given solution (particle) solves an optimization problem. Each particle attempts to improve upon the previous best solution by updating its velocity and position based on individual and collective experience. This process is repeated until a stop condition is satisfied, after which the algorithm aims to determine the best parameters (_ and _) and find the weight matrix for each optimal local model.

[0065] When the termination condition is met and the maximum number of rounds is completed, each on-device AI node has an updated local model that incorporates local data as well as neighbor data weighted by optimized neighbor weights. The weight vectors of each on-device AI node, optimized through the PSO algorithm, can be combined with vectors from other nodes to form part of the weight matrix.

[0066] Figure 7 illustrates pseudocode for the PSO algorithm for deriving the peer weight matrix, which is the weight matrix for the local model as described above.

[0067] Meanwhile, the peer weighting algorithm of the present invention may also use an HMM algorithm.

[0068] Referring to FIGS. 8 to 10, the following is explained: First, as shown in FIG. 8, the input consists of the accuracy (P) and waiting time (W) derived from the performance, similar to the PSO algorithm. Then, the HMM is trained on the observation to find the conversion probability.

[0069] Transition probabilities refer to the probability of transitioning from one hidden state to another. Peer weights can be calculated from these transition probabilities when one state (node) transitions to another state (another node). Peer weights are assigned based on the results and evaluations of the HMM, and this assignment may be probabilistic, considering the probabilities or possibilities derived from the HMM. This may mean that higher peer weights are assigned to states or tasks that the model predicts will yield better performance or shorter latency.

[0070] A flowchart of the process using this HMM algorithm is shown in FIG. 9, and the pseudocode is shown in FIG. 10.

[0071] As can be seen from the diagram, the initial stage can begin by initializing the peer weight matrix R, the observation matrix O, and the state distribution θ. Forward α and inverse β probabilities associated with the state of the HMM are calculated, and these probabilities can inform updates to the observation matrix.

[0072] A check is performed to verify whether the stop condition is satisfied, and if it is not satisfied, the algorithm proceeds iteratively. Within each iteration, the algorithm can compute peer weight updates for the temporary variables γ, ξ, and R, which are used to update the observation matrix O and the state distribution θ. These updates can be performed based on the observed data by re-estimating model parameters based on forward and backward probabilities, neighbor weights, and forward and backward probabilities.

[0073] When the termination condition is satisfied and the maximum round is completed, each on-device AI node (100) can be completed with each local model updated based on its own local data and data obtained from neighboring nodes. The weight vector can be optimized for a specific round and stored in shared memory, and when the maximum round is completed, each on-device AI node (100) can form a peer weight matrix R.

[0074] Referring again to FIG. 4, when the model evaluation unit (20) of the on-device AI node (100) derives weights for each local model, i.e., a peer weight matrix, according to the PSO algorithm or HMM algorithm, the topology management unit (30) then determines the optimal topology using the peer weight matrix derived according to the evaluation results.

[0075] At this time, as described above, when the on-device AI node (100) of the present invention operates as a leader node, it includes a topology management unit (30) and can directly determine the optimal topology, and when it is not a leader node, the on-device AI node (100) transmits a peer weight matrix derived according to the evaluation result to the leader node, and the leader node can determine the optimal topology by aggregating the peer weight matrix transmitted from other nodes.

[0076] The optimal topology defined in the present invention refers to a topology in which a straggler node capable of degrading performance is removed from a plurality of on-device AI nodes (100) interconnected according to one topology in an on-device AI network system (500), as illustrated in FIG. 11. To this end, the topology management unit (30) of the present invention identifies a straggler node using the results of performing an evaluation of a local model and determines an optimal topology in which the identified straggler node is removed from the topology.

[0077] This will be explained with reference to FIGS. 12 to 14.

[0078] First, FIG. 12 is an example diagram for explaining the process of determining the optimal topology performed at the leader node (DTN) of the present invention. As described above, when a plurality of on-device AI nodes (node1, node2, node3, …, node n) constituting an on-device AI network system are interconnected according to one topology (1), each on-device AI node learns a local model using each local data collected in one round (2) and transmits it to at least one connected neighbor node to mutually share the local model, and then evaluates the learned local model and the shared local model (4), applies the latency and the performance of each local model to a peer weighting algorithm to derive a peer weighting vector (or reward vector), which is then transmitted to the leader node (DTN), and the leader node (DTN) can aggregate this.

[0079] Then, the lead node (DTN) applies a set of peer weight vectors, which are node-specific features, i.e., a peer weight matrix, and information about the connection structure, i.e., an adjacency matrix for the topology, to the graph neural network (GNN) to identify lagging nodes (6), and saves the topology with the lagging nodes removed to update the adjacency matrix (7).

[0080] This process is illustrated in more detail in FIG. 13, and the graph neural network (GNN) of the present invention is described more specifically with reference to FIG. 14. The GNN is a neural network that uses data of a graph structure as input. The GNN can be placed in a coordinator to easily receive peer weight vectors through the shared memory of a leader node. It takes a peer weight matrix, which is a characteristic of a node, and an adjacency matrix, which is an initial topology connection structure, as input values, and the output value is a neural network that aims to output information about an optimal network topology based on node functions and resources, that is, an output of a topology weight matrix.

[0081] In this case, the GNN may include two linear layers (linear layer1, 2) that use ReLU activation.

[0082] Subsequently, once the topology weight matrix is ​​obtained through the GNN, nodes with latency exceeding a certain threshold or loss-based performance below a certain threshold are identified as straggler nodes; after determining the optimal topology by removing the straggler nodes, the connections between multiple on-device AI nodes can be reconfigured according to the determined optimal topology.

[0083] Furthermore, the aforementioned process can be carried out until all rounds of distributed federated learning are completed, and by identifying and removing straggler nodes round by round to configure an optimal topology, the robustness, efficiency, and speed of distributed federated learning can be improved.

[0084] Furthermore, although it is difficult to configure dynamic network topologies due to the heterogeneity of on-device AI nodes, distributed federated learning topologies can be dynamically optimized through federated graph representation learning. This enables the optimization of distributed federated learning topologies that minimize divergence of local training models, facilitate rapid training model convergence with non-IID data characteristics, and generalize global models via leader nodes.

[0085] For the above, an on-device AI network system for performing distributed federated learning and a method for configuring a distributed federated learning topology according to an embodiment of the present invention have been described.

[0086] The overall flowchart of the distributed federated learning topology configuration method described above can be seen in Fig. 15.

[0087] The distributed federated learning topology configuration method of the present invention as described above may be provided in the form of a computer-readable medium suitable for storing computer program instructions and data.

[0088] In particular, the computer program of the present invention may execute a first step of learning a local model using each local data collected in one round; a second step of sharing the local model by transmitting the learned local model to at least one connected neighbor node; a third step of evaluating the learned local model and the shared local model; a fourth step of identifying a straggler node using the result of the evaluation; a fifth step of determining an optimal topology by removing the identified straggler node; and a sixth step of reconstructing the mutual connections according to the determined optimal topology.

[0089] Computer-readable media suitable for storing such computer program instructions and data include, for example, recording media such as magnetic media (e.g., hard disks, floppy disks, and magnetic tapes), optical recording media (e.g., CD-ROM, Digital Video Disk), and floptical disks (e.g., magneto-optical media), and semiconductor memories such as ROM (Read Only Memory), RAM (Random Access Memory), flash memory, EPROM (Erasable Programmable ROM), and EEPROM (Electrically Erasable Programmable ROM). Processors and memory may be supplemented by or integrated with special-purpose logic circuits.

[0090] Furthermore, computer-readable recording media may be distributed across networked computer systems, allowing computer-readable code to be stored and executed in a distributed manner. Additionally, the functional program for implementing the present invention, along with related code and code segments, may be easily inferred or modified by programmers skilled in the art to which the present invention pertains, taking into account the system environment of the computer that reads the recording media to execute the program.

[0091] In addition, a computer program recorded on a computer-readable recording medium as described above includes instructions that perform the functions described above, and can execute the aforementioned functions by being distributed and circulated through the recording medium, read, installed, and executed on a specific device or specific computer.

[0092] The present invention has been described with reference to embodiments illustrated in the drawings, but this is merely illustrative, and those skilled in the art will understand that various modifications and equivalent alternative embodiments are possible therefrom. Accordingly, the true technical scope of protection of the present invention should be determined by the technical concept of the claims.

Claims

1. A method for configuring a decentralized federated learning topology in an on-device AI network system comprising multiple on-device AI nodes interconnected according to a single topology and performing decentralized federated learning in rounds, Each of the above plurality of on-device AI nodes is Step 1: Training a local model using each local data collected in one round; A second step of mutually sharing local models by transmitting the learned local model to at least one connected neighbor node; Step 3: Evaluating the learned local model and the shared local model; A fourth step of identifying strggler nodes using the results of the above evaluation; Step 5, determining the optimal topology by removing the identified straggler nodes; and Step 6, reconfiguring the interconnections according to the optimal topology determined above; A method for configuring a distributed federated learning topology in an on-device AI network system, characterized by including 2. In Paragraph 1, The above third step A method for configuring a distributed federated learning topology in an on-device AI network system, characterized by inputting accuracy derived from waiting time and loss-based performance for each local model into a peer weighting algorithm to derive an output peer weighting matrix as an evaluation result.

3. In Paragraph 2, The above peer weighting algorithm is A method for constructing a distributed federated learning topology in an on-device AI network system, characterized by being either Particle Swarm Optimization (PSO) or a Hidden Markov Model (HMM).

4. In Paragraph 1, The above fourth step A method for configuring a distributed federated learning topology in an on-device AI network system characterized by identifying straggler nodes using a Graph Neural Network (GNN).

5. In Paragraph 4, The above fourth step The peer weight matrix derived from the result of performing the above evaluation and the adjacency matrix for the topology in which the plurality of on-device AI nodes are interconnected are input into a graph neural network, and A topology weight matrix for a plurality of on-device AI nodes is obtained from the above graph neural network, and A method for configuring a distributed federated learning topology in an on-device AI network system, characterized by identifying nodes as straggler nodes in the acquired topology weight matrix where the waiting time is above a certain threshold or the loss-based performance is below a certain threshold.

6. In Paragraph 1, After the above Step 6, A method for configuring a distributed federated learning topology in an on-device AI network system, characterized in that, until a designated learning round is completed, each of the multiple on-device AI nodes interconnected and reconfigured according to the determined optimal topology repeats steps 1 through 6.

7. A computer-readable recording medium storing a program for executing a distributed federated learning topology configuration method in an on-device AI network system as described in any one of claims 1 through 6.

8. An on-device AI network system comprising multiple on-device AI nodes interconnected according to a single topology and performing decentralized federated learning in rounds, The above-mentioned on-device AI node is A shared memory storing a local model trained using local data collected in one round and a shared local model transmitted from at least one connected neighbor node; A model evaluation unit that evaluates the above-mentioned learned local model and shared local model; and A topology management unit that identifies straggler nodes using the results of an evaluation performed by the above-mentioned model evaluation unit, and reconstructs mutual connections according to an optimal topology in which the identified straggler nodes are removed; An on-device AI network system characterized by including 9. In Paragraph 8, The above model evaluation department An on-device AI network system characterized by inputting accuracy derived from waiting time and loss-based performance for each local model into a peer weighting algorithm to derive an output peer weighting matrix as an evaluation result.

10. In Paragraph 8, The above topology management department An on-device AI network system characterized by inputting a peer weight matrix derived from the result of performing the above evaluation and an adjacency matrix for a topology in which a plurality of on-device AI nodes are connected into a graph neural network, obtaining a topology weight matrix for each on-device AI node from the graph neural network, and identifying nodes as straggler nodes in the obtained topology weight matrix where the waiting time is above a certain standard or the loss-based performance is below a certain standard.

Images