System and method for constructing learning model

Abstract

A system for constructing a learning model according to the present invention comprises: a server in which a global model is stored and which updates the global model; and a plurality of agent devices which receive the global model from the server, connect the global model between layers of local models, input a plurality of pieces of training data to the local models, and calculate the losses between the label values of the training data and the outputs of the local model and the global model to update the local models, thereby training the local models, wherein the server receives the trained local models from the plurality of agent devices and updates the global model.

Description

Learning model building system and method

[0001] The present invention relates to a learning model construction system and method for constructing a learning model through machine learning.

[0002] Federated learning is a privacy-preserving deep learning model training method. Training is conducted using multiple agents that train models using their own training data and a single server that aggregates the trained models. Federated learning is useful in environments with high demands for data security and privacy. Because it decentralizes data processing rather than centralizing it, it mitigates privacy and security concerns while allowing for the sharing of useful data information to improve learning models.

[0003] However, in conventional federated learning, because the classification level of the training data owned by each agent varies, the local model learning may proceed in different directions for each agent during the local learning process. This situation leads to a mismatch in the learning direction of the local and global models, resulting in reduced learning efficiency and generalization performance.

[0004] Therefore, to overcome this, a technique is required that can learn new knowledge from heterogeneous training data without biasing the representation of the local model through local learning.

[0005] The present invention, in order to solve the above-mentioned problem, provides a learning model construction system and method that provides federated learning that matches the representations of the local model and the global model through knowledge distillation by forming a main path for the local model and a hybrid path that mixes the local model and the global model.

[0006] However, the technical tasks that this embodiment seeks to accomplish are not limited to the technical tasks described above, and other technical tasks may exist.

[0007] As a technical means for solving the above-described technical problem, a learning model construction system according to an embodiment of the present invention includes: a server in which a global model is stored and which updates the global model; and a plurality of agent devices which receive the global model from the server, connect the global model between layers of a local model, input a plurality of learning data into the local model, and calculate a loss between a label value of the learning data and an output of the local model and the global model to update the local model to train the local model, wherein the server receives a local model whose training has been completed from the plurality of agent devices and updates the global model.

[0008] In addition, a method for building a learning model according to another embodiment of the present invention includes a step in which a server transmits a global model to a plurality of agent devices; a step in which the agent devices connect the global model between layers of a local model and input a plurality of learning data into the local model; a step in which the agent devices calculate a loss between a label value of the learning data and an output of the local model and the global model to update the local model; and a step in which the server receives a local model in which learning has been completed from the plurality of agent devices and updates the global model.

[0009] According to the problem solving means of the present invention described above, by applying a knowledge distillation approach through a hybrid path, federated learning can be performed to match the representation of the local model and the global model without limiting the representation of the local model.

[0010] In addition, it can be efficiently applied to federated learning applications involving agents with limited network bandwidth, as it does not incur additional communication costs and does not require the transmission of global information such as global gradient information.

[0011] Additionally, since the agent does not need to store the history information of the local model, the resource usage of the agent can be reduced and the complexity can be reduced, which can increase the participation rate of the agent.

[0012] Figure 1 is a conceptual diagram of a learning model construction system according to one embodiment of the present invention.

[0013] Figure 2 is an example diagram illustrating the operation of training a local model of an agent device.

[0014] Figure 3 is a flowchart illustrating a method for building a learning model according to one embodiment of the present invention.

[0015] Figure 4 is a flowchart for explaining the process of inputting multiple learning data shown in Figure 3 into a local model.

[0016] Hereinafter, the present invention will be described in detail with reference to the attached drawings. However, the present invention can be implemented in various different forms and is not limited to the embodiments described herein. In addition, the attached drawings are only intended to facilitate easy understanding of the embodiments disclosed in the present specification, and the technical ideas disclosed in the present specification are not limited by the attached drawings. In order to clearly explain the present invention in the drawings, parts that are not related to the description are omitted, and the size, shape, and shape of each component shown in the drawings can be variously modified. The same / similar drawing reference numerals are assigned to the same / similar parts throughout the specification.

[0017] The suffixes "module" and "part" used in the following description for components are assigned or used interchangeably solely for the convenience of writing the specification, and do not in themselves have distinct meanings or roles. Furthermore, in describing the embodiments disclosed herein, detailed descriptions of related known technologies have been omitted if they are deemed to obscure the gist of the embodiments disclosed herein.

[0018] Throughout the specification, when a part is said to be "connected (connected, in contact with, or coupled)" to another part, this includes not only cases where it is "directly connected (connected, in contact with, or coupled)" but also cases where it is "indirectly connected (connected, in contact with, or coupled)" with another member in between. Furthermore, when a part is said to "include (have or provide)" a certain component, this does not mean that it excludes other components, but rather that it may "include (have or provide)" other components, unless otherwise specifically stated.

[0019] As used herein, ordinal terms such as "first," "second," etc., are used solely to distinguish one component from another and do not limit the order or relationship of the components. For example, the first component of the present invention may be referred to as the "second component," and similarly, the second component may also be referred to as the "first component."

[0020] Figure 1 is a conceptual diagram schematically illustrating a learning model construction system according to one embodiment of the present invention.

[0021] Referring to FIG. 1, a learning model construction system (1000) according to one embodiment of the present invention is described. The learning model construction system (1000) includes a server (100) and multiple agent devices (200) to construct a learning model through federated learning.

[0022] The server (100) stores the global model to be constructed and transmits the global model to multiple agent devices (200). Thereafter, the learned local model is received from each of the multiple agent devices (200) and the global model is updated.

[0023] The server (100) initially receives the global iteration count, the local iteration count, and the global model, and transmits the global model to some agent devices (200) extracted through random sampling from among a plurality of agent devices (200). In addition, the server (100) transmits the set local iteration count information together with the global model to set the number of local model iteration learning in the agent devices (200). Then, the server receives the local models for which learning has been completed from the plurality of agent devices (200), updates the global model, and repeats the operation of transmitting the updated global model to some agent devices (200) extracted through new random sampling as many times as the global iteration count.

[0024] To perform such operations, the server (100) may be implemented as a computing device including a memory, a processor, and a communication module. The memory stores a global model construction program, and the memory should be interpreted as a general term for non-volatile storage devices that maintain stored information even when power is not supplied, and volatile storage devices that require power to maintain the stored information. The memory may perform the function of temporarily or permanently storing data processed by the processor. In addition to volatile storage devices that require power to maintain stored information, the memory may include magnetic storage media or flash storage media, but the scope of the present invention is not limited thereto.

[0025] And, the processor executes the global model construction program stored in the memory to perform the overall operation for training the global model described above. In this embodiment, the processor may be implemented in the form of a microprocessor, a central processing unit (CPU), a processor core, a multiprocessor, an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA), etc., but the scope of the present invention is not limited thereto. In addition, the processor provides a function for controlling the hardware of the server (100) according to the execution of the global model construction program. That is, the processor may perform hardware control functions such as a necessary file system, memory allocation, network, basic library, timer, device control (display, media, input device, 3D, etc.), and other utilities according to the execution of the program.

[0026] Meanwhile, the server (100) may operate in a cloud computing service model such as Software as a Service (SaaS), Platform as a Service (PaaS), or Infrastructure as a Service (IaaS). Furthermore, the server (100) may be constructed in the form of a private cloud, a public cloud, or a hybrid cloud.

[0027] The agent device (200) receives information on the global model and the number of local iterations from the server (100), connects the global model between layers of the local model, and trains the local model using multiple training data. Specifically, the local model is updated by calculating the loss between the label values ​​of the training data and the outputs of the local model and the global model, and this process is repeated for the number of local iterations to train the local model. Here, the multiple training data may be labeled data stored in the agent device (200). Additionally, the agent device (200) may update the local model by calculating the output loss of the local model and the global model.

[0028] To perform such operations, the agent device (200) may be implemented as a computing device including a memory, a processor, and a communication module. The memory stores a local model building program, and the memory should be interpreted as a general term for non-volatile storage devices that maintain stored information even when power is not supplied, and volatile storage devices that require power to maintain the stored information. The memory may perform the function of temporarily or permanently storing data processed by the processor. In addition to volatile storage devices that require power to maintain stored information, the memory may include magnetic storage media or flash storage media, but the scope of the present invention is not limited thereto.

[0029] And, the processor executes the local model building program stored in the memory to perform the overall operation for training the local model described above. In this embodiment, the processor may be implemented in the form of a microprocessor, a central processing unit (CPU), a processor core, a multiprocessor, an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA), etc., but the scope of the present invention is not limited thereto. In addition, the processor provides a function for controlling the hardware of the agent device (200) according to the execution of the global model building program. That is, the processor can perform hardware control functions such as a necessary file system, memory allocation, network, basic library, timer, device control (display, media, input device, 3D, etc.), and other utilities according to the execution of the program.

[0030] Figure 2 is an exemplary diagram illustrating the operation of training a local model of an agent device. Referring to Figure 2, the operation of training a local model of an agent device (200) will be described in detail.

[0031] The agent device (200) divides the local model into a plurality of local blocks (10) and connects the plurality of local blocks (10) to form a main path. In the main path, the plurality of local blocks (10) form each layer. Referring to Fig. 2, the agent device (200) divides the local model into a first local block (10-1), a second local block (10-2), and a third local block (10-3), and forms a main path by sequentially connecting the first local block (10-1), the second local block (10-2), and the third local block (10-3).

[0032] And, the agent device (200) divides the global model into multiple global models (20) and connects at least one global block (20) to the main path to form multiple hybrid paths. In the hybrid path, at least one global block (20) is connected between the nth (n is a natural number) local block (10) and the n+1th local block (10), and the first global block (20) is formed as the n+1th layer. And, the first global block (20) of each hybrid path calculates a feature value using the feature value of the nth local block (10).

[0033] Referring to FIG. 2, the agent device (200) divides the global model into a first global block (20-1), a second global block (20-2), and a third global block (20-3). Then, the first global block (20-1) and the second global block (20-2) are connected between the first local block (10-1) and the second local block (10-2) to form a first hybrid path, and the third global block (20-3) is connected between the second local block (10-2) and the third local block (10-3) to form a second hybrid path.

[0034] Here, the first hybrid path sequentially forms layers with the first local block (10-1), the first global block (20-1), and the second global block (20-2), and the second hybrid path sequentially forms layers with the first local block (10-1), the second local block (10-2), and the third global block (20-3). And, in the first hybrid path, the first global block (20-1) sequentially forms layers with the first feature value ( ) is received and the feature value is calculated, and in the second hybrid path, the third global block (20-3) calculates the second feature value ( of the second local block (10-2) ) is passed and the feature values ​​are calculated.

[0035] The agent device (200) inputs multiple learning data into the main path formed as described above, and trains local blocks through the main path and multiple hybrid paths. Since multiple hybrid paths are connected to the main path, inputting multiple learning data into the main path can also input multiple learning data into each of the multiple hybrid paths.

[0036] The agent device (200) updates the local model by calculating the label value of the learning data and the output loss between the main path and multiple hybrid paths and the output loss between the main path and multiple hybrid paths.

[0037] Referring to FIG. 2, the agent device (200) outputs a first output value for the main path ( ) and the first loss value between the label value (30) of the training data ( ), multiple second output values ​​for each hybrid path ( , ) and the second loss value between the label value (30) of the training data ( ) and the first output value ( ) and multiple second output values ​​( , ) between the third loss values ​​( ) and calculate the first loss value ( ), the second loss value ( ) and the third loss value ( ) is used to update the local model. Here, the first loss value ( ) and the second loss value ( ) is calculated using cross-entropy. The first loss value ( ) is calculated using mathematical formula 1, and the second loss value ( ) is calculated using mathematical formula 2.

[0038] [Mathematical Formula 1]

[0039]

[0040] [Equation 2]

[0041]

[0042] In mathematical expressions 1 and 2, y is the label value of the training data.

[0043] Also, the third loss value ( ) is calculated using the Kullback-Leibler divergence and is calculated through mathematical expression 3.

[0044] [Equation 3]

[0045]

[0046] The agent device (200) calculates the first loss value ( ), the second loss value ( ) and the third loss value ( ) is used to calculate the parameter gradient, and the multiple local blocks (10-1, 10-2, 10-3) included in the main path are updated using the parameter gradient. The agent device (200) applies the final loss function defined by Equation 4 to the main path, the first hybrid path, and the second hybrid path using the gradient descent method to calculate the parameter gradient for the final loss function, and updates only the parameters included in the main path using the calculated parameter gradient.

[0047] Here, the agent device (200) maintains the parameters for the global block (20) without updating them, and repeats the operation of updating the local model by the number of local iterations. In this process, a knowledge distillation approach is applied, whereby the local block is updated in the learning direction of the global model through the output values ​​of multiple hybrid paths, enabling learning to be performed.

[0048] [Equation 4]

[0049]

[0050] In mathematical formula 4 , is a hyperparameter that determines the weight of each item.

[0051] Thereafter, when the learning of the local model is completed, the agent device (200) transmits the local model whose learning has been completed to the server (100), and the server (100) calculates the average of the parameters of the local model whose learning has been completed from the agent device (100) to update the global model. The server (100) updates the global model using the local models whose learning has been completed received from a plurality of agent devices (200), and transmits the updated global model to some agent devices (200) extracted through new random sampling, thereby training the global model by repeating the operation as many times as the global repetition number.

[0052] Figure 3 is a flowchart illustrating a method for constructing a learning model according to one embodiment of the present invention.

[0053] Hereinafter, a method (S100) for constructing a learning model using a learning model constructing system (1000) according to an embodiment of the present invention will be described with reference to FIGS. 1 to 3. The server (100) receives learning information about an initial global model, a global iteration count, and a local iteration count (step S110), and transmits the global model to a plurality of agent devices (200) (step S120). Each agent device (200) that has received the global model connects the global model between layers of a local model, and then inputs a plurality of learning data into the local model (step S130). Thereafter, the agent device (200) calculates a loss between the label value of the learning data and the output of the local model and the global model to update the local model (step S140), and when learning of the local model is completed in the agent device (200), the server (100) receives the local model whose learning has been completed from the plurality of agent devices (200) and updates the global model (step S150).

[0054] Afterwards, each step of the learning model construction method (S100) is described in detail.

[0055] When describing the process of transmitting a global model from a server (100) to multiple agent devices (200) (step S120), the server (100) transmits the global model to some agent devices (200) selected through random sampling from among the multiple agent devices (200). Here, the server (100) transmits information on the set local repetition count together with the global model to set the number of local model repetition learning in the agent devices (200).

[0056] Figure 4 is a flowchart for explaining the process of training the local model and global model shown in Figure 3.

[0057] Next, with reference to FIGS. 2 and 4, a process (step S130) in which the agent device (200) inputs multiple learning data into the local model is described.

[0058] The agent device (200) divides the local model into a plurality of local blocks (10) and the global model into a plurality of global models (20) (step S131), and connects the plurality of local blocks (10) to form a main path (step S132). In the main path, the plurality of local blocks (10) form each layer. Referring to FIG. 2, the agent device (200) divides the local model into a first local block (10-1), a second local block (10-2), and a third local block (10-3), and forms a main path by sequentially connecting the first local block (10-1), the second local block (10-2), and the third local block (10-3).

[0059] And, at least one global block (20) is connected to the main path to form multiple hybrid paths (step S133). In the hybrid path, at least one global block (20) is connected between the nth (n is a natural number) local block (10) and the n+1th local block (10), and the first global block (20) is formed as the n+1th layer. And, the first global block (20) of each hybrid path calculates a feature value using the feature value of the nth local block (10).

[0060] Referring to FIG. 2, the agent device (200) divides the global model into a first global block (20-1), a second global block (20-2), and a third global block (20-3). Then, the first global block (20-1) and the second global block (20-2) are connected between the first local block (10-1) and the second local block (10-2) to form a first hybrid path, and the third global block (20-3) is connected between the second local block (10-2) and the third local block (10-3) to form a second hybrid path.

[0061] Here, the first hybrid path sequentially forms layers with the first local block (10-1), the first global block (20-1), and the second global block (20-2), and the second hybrid path sequentially forms layers with the first local block (10-1), the second local block (10-2), and the third global block (20-3). And, in the first hybrid path, the first global block (20-1) sequentially forms layers with the first feature value ( ) is received and the feature value is calculated, and in the second hybrid path, the third global block (20-3) calculates the second feature value ( of the second local block (10-2) ) is passed and the feature values ​​are calculated.

[0062] Then, the agent device (200) initiates training of a local model by inputting multiple learning data into a main path to which multiple hybrid paths are connected (step S134). Since multiple hybrid paths are connected to the main path, inputting multiple learning data into the main path can also input multiple learning data into each of the multiple hybrid paths.

[0063] Next, referring to FIG. 2, a process (step S140) in which the agent device (200) updates a local model is described. The agent device (200) trains a local block through the main path and multiple hybrid paths formed as described above, and updates the local model by calculating the label value of the training data and the output loss between the main path and the multiple hybrid paths and the output loss between the main path and the multiple hybrid paths.

[0064] Referring to FIG. 2, the agent device (200) outputs a first output value for the main path ( ) and the first loss value between the label value (30) of the training data ( ), multiple second output values ​​for each hybrid path ( , ) and the second loss value between the label value (30) of the training data ( ) and the first output value ( ) and multiple second output values ​​( , ) between the third loss values ​​( ) and calculate the first loss value ( ), the second loss value ( ) and the third loss value ( ) is used to update the local model. Here, the first loss value ( ) and the second loss value ( ) is calculated using cross-entropy. The first loss value ( ) is calculated using mathematical formula 1, and the second loss value ( ) is calculated using mathematical formula 2.

[0065] [Mathematical Formula 1]

[0066]

[0067] [Equation 2]

[0068]

[0069] In mathematical expressions 1 and 2, y is the label value of the training data.

[0070] Also, the third loss value ( ) is calculated using the Kullback-Leibler divergence and is calculated through mathematical expression 3.

[0071] [Equation 3]

[0072]

[0073] The agent device (200) calculates the first loss value ( ), the second loss value ( ) and the third loss value ( ) is used to calculate the parameter gradient, and the multiple local blocks (10-1, 10-2, 10-3) included in the main path are updated using the parameter gradient. The agent device (200) applies the final loss function defined by Equation 4 to the main path, the first hybrid path, and the second hybrid path using the gradient descent method to calculate the parameter gradient for the final loss function, and updates only the parameters included in the main path using the calculated parameter gradient.

[0074] Here, the agent device (200) maintains the parameters for the global block (20) without updating them, and repeats the operation of updating the local model by the number of local iterations. In this process, a knowledge distillation approach is applied, whereby the local block is updated in the learning direction of the global model through the output values ​​of multiple hybrid paths, enabling learning to be performed.

[0075] [Equation 4]

[0076]

[0077] In mathematical formula 4 , is a hyperparameter that determines the weight of each item.

[0078] Next, the process of updating the global model (step S150) will be described. When the learning of the local model is completed, the agent device (200) transmits the learned local model to the server (100), and the server (100) calculates the parameter average of the learned local model from the agent device (100) to update the global model. Then, if the number of updates of the global model does not satisfy the global number of iterations, the server (100) performs step S120 again. The server (100) transmits the global model updated in the previous round to some agent devices (200) extracted through new random sampling. The server (100) returns the global model if the number of updates of the global model satisfies the global number of iterations.

[0079] The present invention may also be implemented in the form of a recording medium containing computer-executable instructions, such as program modules executed by a computer. Computer-readable media may be any available media that can be accessed by a computer, and include both volatile and nonvolatile media, removable and non-removable media. Computer-readable media may also include computer storage media. Computer storage media includes both volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storing information, such as computer-readable instructions, data structures, program modules, or other data.

[0080] Additionally, although the methods and systems of the present invention have been described with respect to specific embodiments, some or all of their components or operations may be implemented using a computer system having a general-purpose hardware architecture.

[0081] Those skilled in the art will appreciate that the present invention can be readily modified into other specific forms without altering the technical spirit or essential characteristics of the present invention based on the above description. Therefore, the embodiments described above should be understood as illustrative in all respects and not restrictive. The scope of the present invention is defined by the following claims, and all changes or modifications derived from the meaning and scope of the claims and their equivalents should be construed as being included within the scope of the present invention.

[0082] The scope of the present invention is indicated by the claims described below rather than the detailed description above, and all changes or modifications derived from the meaning and scope of the claims and their equivalent concepts should be interpreted as being included in the scope of the present invention.

Claims

1. In a system that builds a learning model through federated learning, A server that stores the global model and updates the global model; and A plurality of agent devices that receive the global model from the server, connect the global model between layers of a local model, input a plurality of learning data into the local model, and update the local model by calculating the loss between the label value of the learning data and the output of the local model and the global model to train the local model, The above server, A learning model construction system that receives local models whose learning has been completed from the plurality of agent devices and updates the global model.

2. In paragraph 1, The above agent device, A learning model building system that updates the local model by calculating the output loss of the local model and the global model.

3. In paragraph 1, The above agent device, The local model and the global model are each divided into a plurality of blocks, and a plurality of local blocks of the local model are connected to form a main path, and at least one global block of the global model is connected to the main path to form a plurality of hybrid paths, and the plurality of learning data are input to the main path. The above main path is formed into each layer by the plurality of local blocks, The above hybrid path is, A learning model construction system in which at least one global block is connected between the nth (n is a natural number) local block and the n+1th local block, and the first global block is formed as the n+1th layer.

4. In paragraph 3, A learning model building system in which the first global block of each hybrid path calculates feature values ​​using the feature values ​​of the nth local block.

5. In paragraph 4, The above agent device, A learning model building system that calculates a first loss value between a first output value for the main path and a label value of the learning data, a second loss value between a plurality of second output values ​​for each hybrid path and a label value of the learning data, and a third loss value between the first output value and the plurality of second output values, and updates the local model using the first loss value, the second loss value, and the third loss value.

6. In paragraph 5, The above agent device, A learning model construction system that calculates a parameter gradient using the first loss value, the second loss value, and the third loss value, and updates a plurality of local blocks included in the main path using the parameter gradient.

7. In paragraph 5, The above agent device, The above first loss value and the above second loss value are calculated using cross-entropy, A learning model construction system in which the above third loss value is calculated using Kullback-Leibler divergence.

8. In the method of building a learning model through federated learning, A step in which the server transmits the global model to multiple agent devices; A step in which the agent device connects the global model between layers of the local model and inputs a plurality of learning data into the local model; A step in which the agent device updates the local model by calculating a loss between the label value of the learning data and the output of the local model and the global model; and A method for building a learning model, comprising a step in which the server receives a local model whose learning has been completed from the plurality of agent devices and updates the global model.

9. In paragraph 8, The steps to update the above local model are: A learning model construction method for updating the local model by calculating the output loss of the local model and the global model.

10. In paragraph 8, The step of inputting the above plurality of learning data into the local model is: A step of dividing the local model and the global model into a plurality of blocks; A step of forming a main path by connecting multiple local blocks of the above local model; A step of forming a plurality of hybrid paths by connecting at least one global block for the global model to the main path; and Including a step of inputting the plurality of learning data into the main path, The above main path is formed into each layer by the plurality of local blocks, The above hybrid path is, A method for constructing a learning model, wherein at least one global block is connected between the nth (n is a natural number) local block and the n+1th local block, and the first global block is formed as the n+1th layer.

11. In paragraph 10, In the step of inputting the above plurality of learning data into the main path, A method for constructing a learning model, wherein the first global block of the above hybrid path calculates feature values ​​using the feature values ​​of the nth local block.

12. In paragraph 11, The steps to update the above local model are: A learning model building system that calculates a first loss value between a first output value for the main path and a label value of the learning data, a second loss value between a plurality of second output values ​​for each hybrid path and a label value of the learning data, and a third loss value between the first output value and the plurality of second output values, and updates the local model using the first loss value, the second loss value, and the third loss value.

13. In paragraph 12, The steps to update the above local model are: A learning model construction method, wherein a parameter gradient is calculated using the first loss value, the second loss value, and the third loss value, and a plurality of local blocks included in the main path are updated using the parameter gradient.

14. In paragraph 12, The steps to update the above local model are: The above first loss value and the above second loss value are calculated using cross-entropy, A method for constructing a learning model, wherein the third loss value is calculated using Kullback-Leibler divergence.

Images