Edge computing device and deep learning model optimization method thereof

Abstract

Provided is a deep learning model optimization method performed by an edge computing device. The method comprises the steps of: measuring a similarity between data collected in an installation environment of the edge computing device and training data used for training a deep learning model deployed in the edge computing device; determining whether to proceed with optimization of the deep learning model on the basis of the result of the similarity measurement; generating training data for proceeding with optimization of the deep learning model on the basis of confidence information; training and generating a deep learning model on the basis of the training data; and updating the existing deep learning model applied to the edge computing device with the generated deep learning model.

Description

Edge computing device and method for optimizing deep learning models thereof

[0001] The present invention relates to an edge computing device and a method for optimizing a deep learning model thereof.

[0002] Applying existing large-scale deep learning models directly to edge computing devices presents various limitations, including limited computing resources. To overcome these limitations, existing deep learning models are being applied in lightweight form. When creating and applying lightweight deep learning models, environmental information, such as weather, can serve as useful information for ensuring stable performance.

[0003] For example, there are two ways to collect weather information: predicting based on a deep learning model or utilizing information from a nearby weather station.

[0004] Weather prediction methods utilizing deep learning models attempt to forecast weather using pre-trained models. This approach has the advantage of making existing large-scale deep learning models lightweight enough to run on edge computing devices. However, this approach suffers from data dependency. Deep learning models rely on training data, making it difficult to provide accurate predictions for weather patterns or situations not included in the training data. Furthermore, because available resources are limited in edge computing environments, model lightweighting is essential.

[0005] On the other hand, utilizing local weather information has the advantage of directly utilizing information about the current environment. This method does not require additional data collection and can obtain weather information by leveraging existing weather data. However, this method suffers from accuracy issues due to distance. Specifically, the difference in distance between the surrounding environment and the weather station can reduce forecast accuracy, and terrain, buildings, and natural obstacles can also affect forecast accuracy. Furthermore, local weather information can also face reliability issues.

[0006] An embodiment of the present invention provides an edge computing device and a deep learning model optimization method thereof for automatically adapting to various installation environments to optimize a weather classification deep learning model.

[0007] However, the technical task that this embodiment seeks to achieve is not limited to the technical task described above, and other technical tasks may exist.

[0008] As a technical means for achieving the above-described technical task, a method performed by an edge computing device according to the first aspect of the present invention includes the steps of: measuring a similarity between data collected in an installation environment of the edge computing device and learning data used for learning a deep learning model installed in the edge computing device; determining whether to proceed with optimization of the deep learning model based on the result of the similarity measurement; generating learning data for proceeding with optimization of the deep learning model based on reliability information; learning and generating a deep learning model based on the learning data; and updating an existing deep learning model applied to the edge computing device with the generated deep learning model.

[0009] In some embodiments of the present invention, the deep learning model may be a deep learning model for weather classification.

[0010] In some embodiments of the present invention, the step of measuring the similarity between data collected in the installation environment of the edge computing device and learning data used for learning a deep learning model installed in the edge computing device may include the steps of: inputting an image collected in the installation environment into a weather classification deep learning model to extract a first feature vector; reducing the dimension of the first feature vector and mapping it onto a first feature space; and comparing information of the mapped first feature vector with information of a second feature vector stored on a pre-built feature space database to measure the similarity.

[0011] In some embodiments of the present invention, the step of measuring the similarity between data collected in the installation environment of the edge computing device and training data used for training a deep learning model installed in the edge computing device may include the steps of: inputting an image used for training into a weather classification deep learning model to extract the second feature vector; reducing the dimension of the second feature vector and mapping it onto a second feature space; and storing information on the mapped second feature vector on a feature space database.

[0012] In some embodiments of the present invention, the step of generating learning data for optimizing the deep learning model based on reliability information generates similarity-based correct answer information when the similarity between the information of the first and second feature vectors satisfies a preset condition, and the learning data may be composed of an image collected from the installation environment and the similarity-based correct answer information including a classification class correct answer label and reliability of the image.

[0013] In some embodiments of the present invention, the step of generating learning data for optimizing the deep learning model based on reliability information may include using the correct class of learning data that satisfies a preset condition as the correct answer label, and applying the reliability determined based on the maximum threshold value of relative entropy (KL-Divergence) and the probability similarity of the first and second feature vectors.

[0014] In some embodiments of the present invention, the step of generating learning data for optimizing the deep learning model based on reliability information may include, when the operating conditions of the edge computing device and the cloud server are such that the similarity does not satisfy a preset condition, inputting an image collected in the installation environment into a weather classification deep learning model trained with predetermined large-scale and multiple domain-based learning data in the cloud server to generate a K-dimensional (K is a natural number) feature vector, and comparing the K-dimensional feature vector with the first feature vector to generate correct answer information based on the most similar feature vector.

[0015] In some embodiments of the present invention, the step of learning and generating a deep learning model based on the learning data may learn the deep learning model based on a loss function that applies the reliability to the difference between the correct label and the predicted value of the deep learning model.

[0016] In some embodiments of the present invention, the step of generating learning data for optimizing the deep learning model based on reliability information may further include the steps of setting weather information of the Korea Meteorological Administration measured at a recent distance based on the installation environment as a correct answer label based on the Korea Meteorological Administration information; calculating a reliability based on the Korea Meteorological Administration information by applying distance information and a predetermined weight based on the location information of the installation environment and the location information corresponding to the recent distance; and generating the Korea Meteorological Administration information-based correct answer label and the Korea Meteorological Administration information-based reliability as the Korea Meteorological Administration information-based correct answer information.

[0017] In some embodiments of the present invention, the step of learning and generating a deep learning model based on the learning data may include the step of applying the reliability to the difference between the correct answer label and the predicted value of the deep learning model; the step of applying the reliability based on the weather information to the difference between the correct answer label based on the weather information and the predicted value of the deep learning model; and the step of learning the deep learning model based on a loss function that adds up the results of each application.

[0018] In addition, an edge computing device according to a second aspect of the present invention includes a communication module for collecting data in an installation environment through a predetermined network, a memory for storing a program for learning and generating a deep learning model, and a processor for executing the program stored in the memory. At this time, the processor measures the similarity between the data collected in the installation environment and the learning data used for learning the deep learning model by executing the program, and if it is determined that optimization of the deep learning model is necessary based on the result of the similarity measurement, learning data for optimization of the deep learning model is generated based on reliability information, and after learning and generating a deep learning model based on the learning data, the existing deep learning model is updated with the generated deep learning model.

[0019] In addition, other methods for implementing the present invention, other systems, and computer-readable recording media recording a computer program for executing the above methods may be further provided.

[0020] According to one embodiment of the present invention described above, there is an advantage in that it is possible to support optimized performance for deep learning models in various environments by automatically analyzing the installed environment.

[0021] Additionally, it has the advantage of improving the quality of deep learning models by generating learning data using various information and reliability measures.

[0022] In addition, by creating a learning model that reflects reliability, it is possible to estimate the prediction reliability of an optimized learning model, allowing predictions using deep learning models to be utilized more reliably.

[0023] Additionally, by automatically optimizing the deep learning model installed on existing edge computing devices, it has the advantage of enabling stable weather classification even when performing weather classification in an environment different from the existing learning environment.

[0024] The effects of the present invention are not limited to the effects mentioned above, and other effects not mentioned will be clearly understood by those skilled in the art from the description below.

[0025] Figure 1 is a flowchart of a deep learning model optimization method according to one embodiment of the present invention.

[0026] FIG. 2 is a diagram for explaining the feature space configuration and similarity measurement process in one embodiment of the present invention.

[0027] FIG. 3a and FIG. 3b are diagrams for explaining a process of generating learning data when there is no weather information in one embodiment of the present invention.

[0028] FIG. 4a and FIG. 4b are diagrams for explaining a process of generating learning data when there is weather information in one embodiment of the present invention.

[0029] FIG. 5 is a block diagram of an edge computing device capable of optimizing a deep learning model according to one embodiment of the present invention.

[0030] The advantages and features of the present invention, and the methods for achieving them, will become clearer with reference to the embodiments described in detail below together with the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below and may be implemented in various different forms. These embodiments are provided solely to ensure that the disclosure of the present invention is complete and to fully inform those skilled in the art of the scope of the present invention, and the present invention is defined solely by the scope of the claims.

[0031] The terminology used herein is for the purpose of describing embodiments only and is not intended to limit the present invention. In this specification, the singular also includes the plural unless specifically stated otherwise. As used herein, the terms "comprises" and / or "comprising" do not exclude the presence or addition of one or more other components in addition to the mentioned components. Like reference numerals refer to like components throughout the specification, and "and / or" includes each and any combination of one or more of the mentioned components. Although "first", "second", etc. are used to describe various components, these components are not limited by these terms. These terms are only used to distinguish one component from another. Therefore, it should be understood that a first component mentioned below may also be a second component within the technical spirit of the present invention.

[0032] Unless otherwise defined, all terms (including technical and scientific terms) used herein may be used in their common sense to those skilled in the art to which the present invention pertains. Furthermore, terms defined in commonly used dictionaries are not to be interpreted ideally or excessively unless explicitly and specifically defined otherwise.

[0033] Hereinafter, a method for optimizing a deep learning model performed by an edge computing device (100) according to one embodiment of the present invention will be described with reference to FIGS. 1 to 4b.

[0034] Figure 1 is a flowchart of a deep learning model optimization method according to one embodiment of the present invention.

[0035] A deep learning model optimization method according to one embodiment of the present invention is implemented by including a step (S110) of measuring the similarity between data collected in an installation environment of an edge computing device (100) and learning data used for learning a deep learning model installed in the edge computing device (100), a step (S120) of determining whether to proceed with optimization of the deep learning model based on the result of the similarity measurement, a step (S130) of generating learning data for proceeding with optimization of the deep learning model based on reliability information, a step (S140) of learning and generating a deep learning model based on the learning data, and a step (S150) of updating an existing deep learning model applied to the edge computing device (100) with the generated deep learning model.

[0036] Meanwhile, each step illustrated in FIG. 1 may be understood to be performed by the edge computing device (100) described below, but is not necessarily limited thereto.

[0037] First, the similarity between the environment in which the edge computing device (100) is installed and the data used in the currently installed deep learning model is measured (S110). Next, a decision is made as to whether to proceed with deep learning model optimization based on the similarity measurement results (S120). In this case, the deep learning model in one embodiment of the present invention may be a deep learning model for weather classification (hereinafter, "weather classification deep learning model").

[0038] FIG. 2 is a diagram for explaining the feature space configuration and similarity measurement process in one embodiment of the present invention.

[0039] In one embodiment, the similarity measurement is to measure the similarity between the feature space DB information of the training data used in the deep learning model and the image of the current installation environment.

[0040] To this end, the image (201) previously used for learning is input into a weather classification deep learning model (202) to extract a feature vector. This is referred to as a second feature vector to distinguish it from the feature vector described later. Next, the dimension of the second feature vector is reduced and mapped onto a second feature space (203), and the information of the mapped second feature vector is stored in a feature space database (204). This process can be repeated for all learning images to build a feature space database.

[0041] The feature space database configured in this way is pre-configured in a general computing environment, reduces the load on the edge computing environment, and can reduce the feature space to an arbitrary N-dimensionality considering the edge computing environment, so it can be utilized in various edge computing environments.

[0042] Next, for similarity analysis, the collected images (211) in the installation environment are input into a weather classification deep learning model (212) to extract a first feature vector. Then, the dimension of the first feature vector is reduced and mapped onto a first feature space (213). Then, the similarity can be measured by comparing the information of the mapped first feature vector with the information of the second feature vector stored in a pre-built feature space database (204) (214).

[0043] As an example, the similarity measure can be applied to the Kullback-Leibler (KL) Divergence method, which measures the difference between two probability distributions. The KL method is used to measure the distance or difference between two probability distributions P and Q, with a higher value indicating greater dissimilarity between the two distributions.

[0044] Referring back to Figure 1, next, learning data for optimizing the deep learning model is generated based on reliability information (S130).

[0045] FIG. 3a and FIG. 3b are diagrams for explaining a process of generating learning data when there is no weather information in one embodiment of the present invention.

[0046] In one embodiment, the confidence-based learning data consists of an image (I) collected from an installation environment, a classification class correct label (L) of the image, and a confidence level (W). In this case, the classification class correct label (L) and confidence level (W) of the image correspond to similarity-based correct information.

[0047] [Formula 1]

[0048]

[0049] Specifically, one embodiment of the present invention operates the learning data generation process differently depending on the operating conditions of the edge computing device (100) environment (Fig. 3a) and, in contrast, the operating conditions of the edge computing device (100) and the cloud server environment (Fig. 3b).

[0050] First, referring to Fig. 3a, an installation environment image (I) is collected (S301), and an N-dimensional feature vector (referred to as a first feature vector) is generated using a weather classification deep learning model and a dimensionality reduction method (S302). Next, the similarity between the information of the generated first feature vector and the probability distribution (KL Divergence) of the information of the second feature vector stored in the feature space database is measured (S303). This is as described in Fig. 2.

[0051] Next, it is determined whether the similarity between the information of the first and second feature vectors satisfies a preset condition (S304). If the determination result is satisfied, correct answer information based on the similarity can be generated (S305). At this time, the similarity condition may be a value arbitrarily set in advance by the administrator.

[0052] At this time, the correct answer label [label] is selected from the similarity-based correct answer information. KL ] uses the correct class of the training data that satisfies the similarity condition as is, and the confidence [w KL ] can apply the reliability determined based on the maximum threshold of relative entropy (KL-Divergence) and the probability similarity of the first and second feature vectors, as in Equation 2.

[0053] [Formula 2]

[0054]

[0055]

[0056] At this time, P in Equation 2 max is the maximum threshold of KL Divergence, and P KL is the probability similarity of the first and second feature vectors, and if it exceeds a certain threshold, the confidence w KL can be defined as 0. That is, in one embodiment of the present invention, Equation 2 is a formula that restricts values ​​with a large difference in similarity from being used in learning.

[0057] Afterwards, once the answer sheet generation is complete, the generation of learning data for the corresponding image is completed.

[0058] Next, referring to FIG. 3b, in the case of the operating environment conditions of the edge computing device (100) and the cloud server, the process from the step of collecting the installation environment image to the step of checking whether the similarity between the information of the first and second feature vectors satisfies the preset conditions (S311 to S315) operates in the same manner as FIG. 3a.

[0059] In contrast, if the similarity does not satisfy the preset conditions, the installation environment image is transmitted to the cloud server (S316), and the cloud server inputs the images collected from the installation environment based on a weather classification deep learning model to generate a K-dimensional feature vector (S317). At this time, the weather classification model of the cloud server may be a deep learning model trained with training data configured based on a predetermined large-scale and multiple domains.

[0060] Next, the similarity is measured by comparing the K-dimensional feature vector with the first feature vector (S318), and the correct answer information can be generated (S320) based on the feature vector with the most similar measured similarity (S319). At this time, steps S319 to S320 are identical to steps S304 to S305 in FIG. 3a described above.

[0061] FIGS. 4A and 4B are diagrams for explaining a process for generating learning data when there is weather information in one embodiment of the present invention. At this time, in one embodiment of the present invention, even when there is weather information, the learning data generation process is operated differently depending on the operating conditions of the edge computing device (100) environment (FIG. 4A), as in the embodiments of FIGS. 3A and 3B, and, unlike this, the operating conditions of the edge computing device (100) and the cloud server environment (FIG. 4B).

[0062] First, referring to Fig. 4a, an installation environment image (I) is collected (S401), and an N-dimensional first feature vector is generated using a weather classification deep learning model and a dimensionality reduction method (S402). Next, the similarity between the information of the generated first feature vector and the probability distribution (KL Divergence) of the information of the second feature vector stored in the feature space database is measured (S403). Next, it is determined whether the similarity between the information of the first and second feature vectors satisfies a preset condition (S404), and if the determination result is satisfied, correct answer information based on the similarity can be generated (S405). These steps S401 to S405 are identical to the steps S301 to S305 described above.

[0063] In addition, one embodiment of the present invention provides weather information from the Korea Meteorological Administration measured at the closest distance based on the installation environment, and a correct answer label based on the Korea Meteorological Administration information [label location ] and the distance information D based on the location information of the installation environment and the location information corresponding to the most recent distance and the predetermined weight w D Reliability based on weather information by applying [w Location ] can be produced (S406). At this time, the distance information D is a function that measures the distance based on latitude and longitude, and lat p1 ,long p1 It represents the latitude and longitude measured by the Korea Meteorological Administration information, and lat p2 ,long p2 Indicates the latitude and longitude in the installation environment.

[0064] [Formula 3]

[0065]

[0066] Next, referring to FIG. 4b, in the case of the operating environment conditions of the edge computing device (100) and the cloud server, the process from the step of collecting the installation environment image to the step of generating the correct answer information based on the weather information (S411 to S416) operates in the same manner as FIG. 4a.

[0067] In addition, if the preset similarity condition is not satisfied, the process from step (S417) to step S421) of generating similarity-based correct answer information using the weather classification deep learning model of the cloud server is the same as in Fig. 3b, and if the weather information exists, a process of generating correct answer information based on the weather information is added in step S422.

[0068] Referring back to Figure 1, when the generation of learning data is completed, a deep learning model is trained and generated based on the generated learning data (S140).

[0069] In step S140, N images are collected by repeating the preceding step S130, and the collected training data can be used to train a weather classification deep learning model. At this time, the reliability of the generated training data can be reflected and trained using both the collected training data L and W.

[0070] Meanwhile, the loss functions used in the process of training a deep learning model using training data are as shown in Equations 4 and 5. Here, Equation 4 is the loss function in the absence of weather information, and Equation 5 is the loss function in the presence of weather information.

[0071] First, referring to equation 4, if there is no weather information, the correct label is label KL and the predicted value p of the deep learning model predict A deep learning model can be trained through a loss function (Loss) that is calculated by applying confidence to the difference between the two.

[0072] [Formula 4]

[0073]

[0074] Referring to equation 5, if there is weather information, the correct label is label KL and the predicted value p of the deep learning model predict Reliability of the difference between w KL Apply and label the correct answer based on the weather information location and the predicted value p of the deep learning model predict Reliability of weather information based on the difference between location After applying, the deep learning model can be trained through a loss function (Loss) that adds up the results of each application.

[0075] [Formula 5]

[0076]

[0077] When the training of the deep learning model is completed, the existing deep learning model applied to the edge computing device (100) is updated with the generated deep learning model (S150).

[0078] Meanwhile, in the above description, steps S110 to S422 may be further divided into additional steps or combined into fewer steps, depending on the implementation of the present invention. Furthermore, some steps may be omitted as needed, and the order of the steps may be changed. Furthermore, even if other omitted details are present, the details described in FIGS. 1 to 4b also apply to the edge computing device (100) of FIG. 5.

[0079]

[0080] FIG. 5 is a block diagram of an edge computing device (100) capable of optimizing a deep learning model according to one embodiment of the present invention.

[0081] An edge computing device (100) according to one embodiment of the present invention includes a communication module (110), a memory (120), and a processor (130).

[0082] The communication module (110) collects data in the installation environment through a predetermined network. Such a communication module (110) may include both a wired communication module and a wireless communication module. The wired communication module may be implemented as a power line communication device, a telephone line communication device, a cable home (MoCA), Ethernet, IEEE1294, an integrated wired home network, and an RS-485 control device. In addition, the wireless communication module may be configured as a module for implementing functions such as WLAN (wireless LAN), Bluetooth, HDR WPAN, UWB, ZigBee, Impulse Radio, 60GHz WPAN, Binary-CDMA, wireless USB technology, wireless HDMI technology, and other 5G (5th generation communication), LTE-A (long term evolution-advanced), LTE (long term evolution), and Wi-Fi (wireless fidelity).

[0083] A program for learning and generating a deep learning model based on data is stored in the memory (120), and the processor (130) executes the program stored in the memory (120). Here, the memory (120) is a general term for a non-volatile storage device and a volatile storage device that maintain stored information even when power is not supplied.

[0084] For example, the memory (120) may include NAND flash memory such as a compact flash (CF) card, a secure digital (SD) card, a memory stick, a solid-state drive (SSD), and a micro SD card, a magnetic computer storage device such as a hard disk drive (HDD), and an optical disc drive such as a CD-ROM or DVD-ROM.

[0085] The processor (130) measures the similarity between data collected in the installation environment and learning data used for learning a deep learning model by executing a program stored in the memory (120), and if it is determined that optimization of the deep learning model is necessary based on the similarity measurement result, learning data for optimizing the deep learning model is generated based on reliability information. Then, the processor (130) learns and generates a deep learning model based on the learning data, and then updates and optimizes an existing deep learning model with the generated deep learning model.

[0086] The deep learning model optimization method according to one embodiment of the present invention described above can be implemented as a program (or application) and stored in a medium to be executed in combination with a hardware server.

[0087] The above-described program may include codes coded in a computer language, such as C, C++, JAVA, or machine language, that can be read by the processor (CPU) of the computer through the device interface of the computer, so that the computer reads the program and executes the methods implemented as a program. Such codes may include functional codes related to functions that define functions necessary for executing the methods, and may include control codes related to execution procedures necessary for the processor of the computer to execute the functions according to a predetermined procedure. In addition, such codes may further include memory reference-related codes regarding which location (address address) of the internal or external memory of the computer should reference additional information or media necessary for the processor of the computer to execute the functions. In addition, if the processor of the computer needs to communicate with any other computer or server located remotely in order to execute the functions, the code may further include communication-related code regarding how to communicate with any other computer or server located remotely using the communication module of the computer, and what information or media to send and receive during communication.

[0088] The above storage medium refers to a medium that stores data semi-permanently and can be read by a device, rather than a medium that stores data for a short period of time, such as a register, cache, or memory. Specifically, examples of the storage medium include, but are not limited to, ROM, RAM, CD-ROM, magnetic tape, floppy disk, and optical data storage device. That is, the program can be stored in various recording media on various servers that the computer can access or in various recording media on the user's computer. In addition, the medium can be distributed across network-connected computer systems, so that computer-readable code can be stored in a distributed manner.

[0089] The steps of a method or algorithm described in connection with an embodiment of the present invention may be implemented directly in hardware, implemented as a software module executed by hardware, or implemented by a combination thereof. The software module may reside in a random access memory (RAM), a read only memory (ROM), an erasable programmable ROM (EPROM), an electrically erasable programmable ROM (EEPROM), a flash memory, a hard disk, a removable disk, a CD-ROM, or any other form of computer-readable recording medium well known in the art to which the present invention pertains.

[0090] While the embodiments of the present invention have been described above with reference to the attached drawings, those skilled in the art will appreciate that the present invention can be implemented in other specific forms without altering the technical concept or essential features thereof. Therefore, the embodiments described above should be understood to be illustrative in all respects and not restrictive.

Claims

1. A method performed by an edge computing device, A step of measuring the similarity between data collected in the installation environment of an edge computing device and learning data used for learning a deep learning model installed in the edge computing device; A step for determining whether to proceed with optimization of a deep learning model based on the above similarity measurement results; A step of generating learning data for optimizing the above deep learning model based on reliability information; A step of learning and generating a deep learning model based on the above learning data; and A step of updating an existing deep learning model applied to the edge computing device with the generated deep learning model, How to optimize deep learning models.

2. In paragraph 1, The above deep learning model is a deep learning model for weather classification. How to optimize deep learning models.

3. In paragraph 1, The step of measuring the similarity between the data collected in the installation environment of the edge computing device and the learning data used for learning the deep learning model installed in the edge computing device is as follows. A step of inputting the collected images in the above installation environment into a weather classification deep learning model to extract a first feature vector; A step of reducing the dimension of the first feature vector and mapping it onto the first feature space; and Comprising a step of measuring similarity by comparing the information of the first feature vector mapped above with the information of the second feature vector stored in a pre-built feature space database. How to optimize deep learning models.

4. In paragraph 3, The step of measuring the similarity between the data collected in the installation environment of the edge computing device and the learning data used for learning the deep learning model installed in the edge computing device is as follows. A step of extracting the second feature vector by inputting the image used in the above learning into a weather classification deep learning model; A step of reducing the dimension of the second feature vector and mapping it onto the second feature space; and Comprising a step of storing information of the above mapped second feature vector in a feature space database, How to optimize deep learning models.

5. In paragraph 3, The step of generating learning data for optimizing the above deep learning model based on reliability information is as follows. If the similarity between the information of the first and second feature vectors satisfies a preset condition, the correct answer information based on the similarity is generated. The above learning data is composed of images collected in the installation environment and the similarity-based correct answer information including the classification class correct answer label and confidence of the image. How to optimize deep learning models.

6. In paragraph 5, The step of generating learning data for optimizing the above deep learning model based on reliability information is as follows. The above correct answer label uses the correct class of the learning data that satisfies the preset conditions as it is, and the above reliability applies the reliability determined based on the maximum threshold of the relative entropy (KL-Divergence) and the probability similarity of the first and second feature vectors. How to optimize deep learning models.

7. In paragraph 5, The step of generating learning data for optimizing the above deep learning model based on reliability information is, if the operating conditions of the edge computing device and cloud server are If the above similarity does not satisfy the preset condition, the image collected in the installation environment is input into the weather classification deep learning model learned with a predetermined large-scale and multiple domain-based learning data in the cloud server to generate a K-dimensional (K is a natural number) feature vector, and the K-dimensional feature vector is compared with the first feature vector to generate correct answer information based on the most similar feature vector. How to optimize deep learning models.

8. In paragraph 5, The step of learning and generating a deep learning model based on the above learning data is as follows: The deep learning model is trained based on a loss function that applies the confidence to the difference between the correct answer label and the predicted value of the deep learning model. How to optimize deep learning models.

9. In paragraph 5, The step of generating learning data for optimizing the above deep learning model based on reliability information is as follows. A step for setting weather information from the Korea Meteorological Administration measured at a recent distance based on the above installation environment as a correct answer label based on weather information from the Korea Meteorological Administration; A step for calculating reliability based on weather information by applying distance information and a predetermined weight based on location information of the above installation environment and location information corresponding to the recent distance; and Further comprising a step of generating a correct answer label based on the above weather information and a reliability based on the weather information as correct answer information based on the weather information. How to optimize deep learning models.

10. In paragraph 9, The step of learning and generating a deep learning model based on the above learning data is as follows: A step of applying the confidence to the difference between the correct answer label and the predicted value of the deep learning model; A step of applying the reliability based on the Meteorological Agency information to the difference between the correct answer label based on the Meteorological Agency information and the predicted value of the deep learning model; and A step of learning the deep learning model based on a loss function that sums up each application result, How to optimize deep learning models.

11. A communication module that collects data in the installation environment through a specified network. Memory where programs for learning and generating deep learning models are stored and Including a processor that executes a program stored in the above memory, The processor measures the similarity between the data collected in the installation environment and the learning data used for learning the deep learning model by executing the program, and if it is determined that optimization of the deep learning model is necessary based on the result of the similarity measurement, it generates learning data for optimization of the deep learning model based on reliability information, learns and generates a deep learning model based on the learning data, and then updates the existing deep learning model with the generated deep learning model. Edge computing devices.

Images