KR20210068707A - Method for training neural network - Google Patents

Abstract

According to one embodiment of the present disclosure, disclosed is a computer program stored in a computer-readable storage medium. When the computer program is executed on one or more processors to perform the following operations for training a neural network, the operation may comprise: an operation of inputting a first training data set comprising one or more tasks to a first model; an operation of pre-training the first model; and an operation of training a second model comprising at least a part of the pre-trained first model.

Description

How to train a neural network {METHOD FOR TRAINING NEURAL NETWORK}

The present disclosure relates to an information processing method using a computer, and more particularly, to a method for learning a neural network.

As data that can be temporarily or permanently stored and used in a database is accumulated, research on automated data processing of computing devices is being conducted. In order to implement a method for judging the state of data, research on artificial intelligence technology using artificial neural networks is being actively conducted.

In artificial intelligence technology, especially in the field of deep learning, remarkable progress has been made in data automation processing.

In order to train a deep learning model, it is becoming common to learn a neural network by inputting a large amount of data into an artificial neural network. As a result, technology for a method for efficiently training a neural network has been developed.

A training data set is used to train the neural network. In this case, the training data set may be sufficient or insufficient to train the neural network.

When the training data set is sufficient, the technology to train the neural network in a faster time is advancing.

Conversely, when the training data set is insufficient, a technique for acquiring high inference accuracy by training a neural network using the insufficient training data set is also being developed. In this context, the learning method of neural networks is transfer learning method. Transfer learning is one of the neural network learning techniques that reuse a neural network trained on a task rich in training data to build a model of a task lacking training data.

Therefore, the demand for transfer learning methods is increasing in order to solve the problem of insufficient number of training data.

Korean Patent Application Laid-Open No. 10-2019-0098106 discloses a batch normalization layer training method.

The present disclosure has been devised in response to the above-described background technology, and is intended to provide a method for learning a neural network.

According to an embodiment of the present disclosure for realizing the above-described problems, a computer program including instructions stored in a computer-readable storage medium to cause the computer to perform the following operations is disclosed. The operations may include: inputting a first training data set comprising one or more tasks to a first model; pre-training the first model; and training a second model including at least a portion of the pre-trained first model.

In an alternative embodiment, the first training data set may include one or more subsets of the first training data per task.

In an alternative embodiment, the task may include at least one of an image data processing task, a voice data processing task, a natural language processing task, or a time series data processing task as a processing operation using the learned model.

In an alternative embodiment, the image data processing task includes a task of classifying the input image data based on a predetermined criterion, a task of segmenting the input image data, and an anomaly ( It may include at least one of a task for detecting anomaly and a task for recognizing an object from input image data.

In an alternative embodiment, pre-training the first model may include training the first model to perform one or more tasks included in the first training data set using the first model.

In an alternative embodiment, the operation of pre-training the first model may include pre-training the first model in order to minimize the loss of the first model obtained by inputting the first training data set to the first model. It can include actions.

In an alternative embodiment, the operation of pre-training the first model includes: a first training operation of respectively training the first model for each training epoch with respect to the first model; and a second learning operation for learning each of the first models that have undergone one or more first learning operations having different learning epochs.

In an alternative embodiment, the first training operation may include inputting at least a portion of the first training data set to the first model to learn by a predetermined first training epoch.

In an alternative embodiment, the second learning operation may include inputting each of the first training data subsets corresponding to different tasks into the first model subjected to the first learning operation to learn by a predetermined second training epoch. can

In an alternative embodiment, the pre-training of the first model includes: determining the end of learning of the first learning operation based on loss values of one or more first models that have undergone the second learning operation; and determining the first model trained by the determined learning epoch as the pre-trained first model.

In an alternative embodiment, the loss value of the one or more first models that have undergone the second learning operation is a task performance accuracy of each of the one or more first models that have undergone the second learning operation, wherein the second learning operation corresponding to the one or more tasks is selected. Each of the loss values of the coarse first model may be included.

In an alternative embodiment, training the second model including at least a portion of the pre-trained first model may include training the second model by inputting a second training data set to the second model.

In an alternative embodiment, the second training data set is a training data set including a smaller number of data than the first training data set, and may include a training data set corresponding to a single task.

In an alternative embodiment, the second model is a model for performing a single task included in the second training data set, and may include at least one of at least a portion of the pre-trained first model or a regularization layer specialized for a single task. can

In an alternative embodiment, the training of the second model including at least a portion of the pre-trained first model may include fixing weights corresponding to at least a portion of the pre-trained first model and using only a single task-specific regularization layer. It may include a learning operation.

According to another embodiment of the present disclosure, a computer-readable recording medium storing a data structure corresponding to a parameter of a neural network that is at least partially updated in a learning process is disclosed. the operation of the neural network is based at least in part on the parameters, the learning process comprising: inputting a first training data set comprising one or more tasks to a first model; pre-training the first model; and training a second model including at least a portion of the pre-trained first model.

According to another embodiment of the present disclosure, a method for training a neural network is disclosed. The method includes inputting a first training data set comprising one or more tasks to a first model; pre-training the first model; and training a second model including at least a portion of the pre-trained first model.

A computing device for learning a neural network is disclosed according to another embodiment of the present disclosure. one or more processors; and a memory storing instructions executable by the one or more processors; wherein the one or more processors are configured to input a first training data set comprising one or more tasks to a first model, pre-train the first model, and include at least a portion of the pre-trained first model. A second model can be trained.

According to an embodiment of the present disclosure, a computing device for learning a neural network may be provided.

BRIEF DESCRIPTION OF THE DRAWINGS So that the above-mentioned features of the present disclosure may be understood in detail, with a more specific description, with reference to the following embodiments, some of the embodiments are shown in the accompanying drawings. Also, like reference numerals in the drawings are intended to refer to the same or similar functions throughout the various aspects. However, it should be noted that the accompanying drawings only show certain typical embodiments of the present disclosure and are not to be considered as limiting the scope of the present disclosure, and other embodiments having the same effect may be fully appreciated. Take note.
1 is a block diagram of a computing device for learning a neural network according to an embodiment of the present disclosure.
2 is a schematic diagram exemplarily illustrating a neural network in a method for training a neural network, according to an embodiment of the present disclosure.
3 is an exemplary diagram for explaining an operation for pre-training the first model.
4 is an exemplary diagram for explaining an operation for learning a second model.
5 is a flowchart for training a neural network.
6 is a block diagram illustrating a module for learning a neural network.
7 is a simplified, general schematic diagram of an exemplary computing environment in which embodiments of the present disclosure may be implemented.

Various embodiments are now described with reference to the drawings. In this specification, various descriptions are presented to provide an understanding of the present disclosure. However, it is apparent that these embodiments may be practiced without these specific descriptions.

The terms “component,” “module,” “system,” and the like, as used herein, refer to a computer-related entity, hardware, firmware, software, a combination of software and hardware, or execution of software. For example, a component can be, but is not limited to being, a process running on a processor, a processor, an object, a thread of execution, a program, and/or a computer. For example, both an application running on a computing device and the computing device may be a component. One or more components may reside within a processor and/or thread of execution. A component may be localized within one computer. A component may be distributed between two or more computers. In addition, these components can execute from various computer readable media having various data structures stored therein. Components may communicate via a network such as the Internet with another system, for example via a signal having one or more data packets (eg, data and/or signals from one component interacting with another component in a local system, distributed system, etc.) may communicate via local and/or remote processes depending on the data being transmitted).

In addition, the term “or” is intended to mean an inclusive “or” rather than an exclusive “or.” That is, unless otherwise specified or clear from context, "X employs A or B" is intended to mean one of the natural implicit substitutions. That is, X employs A; X employs B; or when X employs both A and B, "X employs A or B" may apply to either of these cases. It should also be understood that the term “and/or” as used herein refers to and includes all possible combinations of one or more of the listed related items.

Also, the terms "comprises" and/or "comprising" should be understood to mean that the feature and/or element in question is present. However, it should be understood that the terms "comprises" and/or "comprising" do not exclude the presence or addition of one or more other features, elements and/or groups thereof. Also, unless otherwise specified or unless the context is clear as to designating a singular form, the singular in the specification and claims should generally be construed to mean "one or more."

In addition, the term “at least one of A or B” should be interpreted as meaning “includes only A”, “includes only B”, and “in the case of a combination of A and B”.

Those skilled in the art will further appreciate that the various illustrative logical blocks, configurations, modules, circuits, means, logics, and algorithm steps described in connection with the embodiments disclosed herein may be implemented in electronic hardware, computer software, or combinations of both. It should be recognized that they can be implemented with To clearly illustrate this interchangeability of hardware and software, various illustrative components, blocks, configurations, means, logics, modules, circuits, and steps have been described above generally in terms of their functionality. Whether such functionality is implemented as hardware or software depends upon the particular application and design constraints imposed on the overall system. Skilled artisans may implement the described functionality in varying ways for each particular application. However, such implementation decisions should not be interpreted as causing a departure from the scope of the present disclosure.

Descriptions of the presented embodiments are provided to enable those skilled in the art to use or practice the present invention. Various modifications to these embodiments will be apparent to those skilled in the art of the present disclosure. The generic principles defined herein may be applied to other embodiments without departing from the scope of the present disclosure. Thus, the present invention is not limited to the embodiments presented herein. This invention is to be construed in its widest scope consistent with the principles and novel features presented herein.

1 is a block diagram of a computing device for learning a neural network according to an embodiment of the present disclosure.

The configuration of the computing device 100 shown in FIG. 1 is only a simplified example. In an embodiment of the present disclosure, the computing device 100 may include other components for performing the computing environment of the computing device 100 , and only some of the disclosed components may configure the computing device 100 .

The computing device 100 may include a processor 110 , a memory 130 , and a network unit 150 .

The processor 110 may include one or more cores, and a central processing unit (CPU) of a computing device, a general purpose graphics processing unit (GPGPU), and a tensor processing unit (TPU). unit) and the like, and may include a processor for data analysis and deep learning. The processor 110 may read a computer program stored in the memory 130 and perform data processing for machine learning according to an embodiment of the present disclosure. According to an embodiment of the present disclosure, the processor 110 may perform an operation for learning the neural network. The processor 110 for learning of the neural network, such as processing input data for learning in deep learning (DL), extracting features from the input data, calculating an error, updating the weight of the neural network using backpropagation calculations can be performed. At least one of a CPU, a GPGPU, and a TPU of the processor 110 may process learning of a network function. For example, the CPU and the GPGPU can process learning of a network function and data classification using the network function together. In addition, in an embodiment of the present disclosure, learning of a network function and data classification using the network function may be processed by using the processors of a plurality of computing devices together. In addition, the computer program executed in the computing device according to an embodiment of the present disclosure may be a CPU, GPGPU, or TPU executable program.

According to an embodiment of the present disclosure, the processor 110 may train a neural network. According to an embodiment of the present disclosure, the processor 110 may train a first model that is a neural network, and then train a second model including at least a portion of the first model. Specifically, after pre-training the first model, which is a neural network, the processor 110 may transfer-learning the second model including at least a part of the first model. The foregoing is merely an example, and the present disclosure is not limited thereto.

According to an embodiment of the present disclosure, the processor 110 may input a first training data set including one or more tasks to the first model.

According to an embodiment of the present disclosure, the first model may include a machine learning model. The first model may include a deep learning model. The first model may include a neural network in which a plurality of nodes are connected. According to an embodiment of the present disclosure, the first model may include a model capable of performing at least one or more tasks. The first model may be a model including information on at least one or more tasks. Accordingly, the first model may be a multi-task complex model capable of performing at least one task. A task may be a group of actions that may be performed using an artificial intelligence model. For example, task 1 may be a task of classifying dogs and cats. Also, task 2 may be a task of classifying lions and elephants. The processor 110 may perform both task 1 and task 2 using the learned first model. Accordingly, when a dog image is input to the learned first model, the processor 110 may output a classification result of dog using the learned first model. Also, when an elephant image is input to the learned first model, the processor 110 may obtain a classification result of an elephant using the learned first model. According to an embodiment of the present disclosure, the first training data set may be a training data set for training the first model. The first training data set may include one or more subsets of the first training data for each task. The foregoing is merely an example, and the present disclosure is not limited thereto.

According to an embodiment of the present disclosure, the first training data set may include one or more subsets of the first training data for each task. The first training data subset may be a training data set corresponding to a specific task. For example, task 1 may be a task of classifying dogs and cats. The first training data subset corresponding to task 1 may be a training data set using dog image data and cat image data as training input data, and dogs and cats as labels, respectively. As another example, task 2 may be a task of classifying elephants and lions. The first training data subset corresponding to task 2 may be a training data set using elephant image data and lion image data as training input data, and using elephant and lion as labels, respectively. Accordingly, the first training data set may include a first training data subset corresponding to task 1 and a second training data subset corresponding to task 2 . Therefore, when the first model is trained by inputting the first training data set including at least one or more tasks to the first model, the processor 110 may obtain a first model capable of performing at least one or more tasks. . The foregoing is merely an example, and the present disclosure is not limited thereto.

According to an embodiment of the present disclosure, the first training data set may include one or more subsets of the first training data for each task. The task is a task processed using the learned model, and may include at least one of an image data processing task, a voice data processing task, a natural language processing task, and a time series data processing task. The above-described task is merely an example, and the present disclosure is not limited thereto.

According to an embodiment of the present disclosure, the image data processing task may include processing by inputting image data into a learned model. The image data processing task includes a task of classifying input image data based on a predetermined criterion, a task of segmenting input image data, a task of detecting anomalies in the input image data, or a task of recognizing an object in the input image data. It may include at least one of the tasks. The task of classifying the image data based on a predetermined criterion may include a task of classifying an object included in the image data. The task of classifying the image data based on a predetermined criterion may include, for example, a task of classifying whether an object included in the image data is a dog or a cat. The task of segmenting the input image data may include a task of classifying pixels included in the image and classifying the boundary of the object in the image from the background to detect the object. The segmentation task may include a semantic segmentation task and/or an instance segmentation task. After performing object segmentation through semantic segmentation, the processor 110 may divide objects of the same class into the same region or color. For example, if there are at least one chair in the image, the processor 110 may divide the chairs into the same area or color area. The processor 110 may classify objects into different objects through instance segmentation, even when objects are of the same type. For example, if there are at least one chair in the image, the processor 110 may divide the chairs into different regions or colors. The task of detecting anomaly in the input image data may include a task of detecting a portion of an image having an irregular pattern that is not normal. For example, the task of detecting anomaly may include a task of detecting a defective part of the product in an image including the product. The task of recognizing an object in the input image data may include a task for recognizing at least one or more objects in the image. The task of recognizing an object from the input image data may include a task of recognizing an object such as each person, ship, building, car, etc. when the image includes an object such as a person, ship, building, car, etc. . The foregoing is merely an example, and the present disclosure is not limited thereto.

According to an embodiment of the present disclosure, the voice data processing task may include an operation that is processed by inputting voice data into a learned model. The voice data processing task may include a task of receiving voice data to perform voice recognition and/or a task of receiving voice data and performing voice generation. The task of performing speech recognition may include a task of converting a human speech language into text data. The task of generating the voice may include a task of outputting voice data having a high correlation with the input voice. For example, the task of performing voice generation includes a task of generating a voice answer to a human question (for example, a task performed by an artificial intelligence speaker capable of talking with a human), composing using artificial intelligence, and performing artificial intelligence. It may include the creation of accompaniment using. The foregoing is merely an example, and the present disclosure is not limited thereto.

According to an embodiment of the present disclosure, the natural language processing task may include an operation processed by inputting a natural language into a learned model. Natural language may include languages that people use on a daily basis. The natural language processing task may include processing by inputting natural language in the form of text data into the learned model. For example, the natural language processing task may include a speaker's intention analysis, a semantic analysis of a sentence, natural language generation, and the like. The foregoing is merely an example, and the present disclosure is not limited thereto.

According to an embodiment of the present disclosure, the time-series data processing task may include processing by inputting time-series data into a learned model. The time series data may include a sequence of data arranged at regular time intervals. The time series data may include, for example, data obtained from a sensor. The time series data processing task may include, for example, predicting stock prices in the financial field, detecting abnormal operation of machinery in the manufacturing industry, and the like. The foregoing is merely an example, and the present disclosure is not limited thereto.

Hereinafter, the operation of pre-training the first model will be described.

According to an embodiment of the present disclosure, the processor 110 may input a first training data set including one or more tasks to the first model. The processor 110 may pre-train the first model using the first training data set. According to an embodiment of the present disclosure, the processor 110 may train the first model to perform one or more tasks included in the first training data set using the first model. When the first model is trained using the first training data set, the processor 110 may obtain a pre-trained first model capable of performing at least one task included in the first training data set. For example, task 1 is a task of detecting anomalies in an image data set obtained from company A, task 2 is a task of detecting anomalies in an image data set obtained from company B, and task 3 is a task of detecting anomalies in an image data set obtained from company B. It can be the task of detecting anomalies in an image data set. The processor 110 may obtain a pre-trained first model using the first training data set. The processor 110 may perform all of the tasks 1, 2, and 3 using the pre-trained first model. The foregoing is merely an example, and the present disclosure is not limited thereto.

According to an embodiment of the present disclosure, the operation of pre-training the first model is an operation of pre-training the first model in order to minimize the loss value of the first model obtained by inputting the first training data set to the first model may include. The loss value may include a value determined based on the difference between the value predicted by the model and the actual value of the label. The loss value may include, for example, a mean square error or a cross entropy error. The processor 110 may minimize the loss value of the first model through the backpropagation method. The backpropagation method may include a method of updating weights included in the model through gradient descent. By learning the first model in a direction that minimizes the loss value, the processor 110 may obtain a model that performs at least one task with high prediction accuracy. The foregoing is merely an example, and the present disclosure is not limited thereto.

According to an embodiment of the present disclosure, the operation of pre-training the first model includes a first learning operation of respectively learning the first model for each learning epoch with respect to the first model, and one or more first learning operations having different learning epochs. It may include a second learning operation of learning each of the first model that has undergone

Hereinafter, a detailed description of the first learning operation among operations of pre-learning the first model will be described with reference to FIG. 3 .

According to an embodiment of the present disclosure, the processor 110 may perform a first learning operation 210 for learning the first model for each training epoch with respect to the first model, respectively. The first learning operation 210 may include an operation of inputting at least a portion of the first training data set into the first model and learning by a predetermined first training epoch. According to an embodiment of the present disclosure, the training epoch inputs the training input data to the input node for all training data included in the training data set, respectively, compares the labeled data with the output of the model to derive an error, By propagating the derived error from an output layer of one or more network functions of the model to an input layer through one or more hidden layers, it may be an operation of updating a weight set in each link. That is, when an operation using a model and a weight update process for the model are performed on all the training data included in the training data set, it may be 1 training epoch. According to an embodiment of the present disclosure, the first predetermined learning epoch may include an epoch determined by an operator and/or an epoch automatically determined by the computing device. The predetermined first learning epoch may include 1 epoch, 2 epochs, 3 epochs, and the like. The foregoing is merely an example, and the present disclosure is not limited thereto.

According to an embodiment of the present disclosure, the first learning operation 210 may include an operation of learning at least one of all tasks included in the first learning data set. The first learning operation 210 may include an operation of respectively learning the first model for each training epoch with respect to the first model. Specifically, in order to improve the learning speed and reduce the amount of computation, the processor 110 may generate a learning data set by combining data sampled from each of the one or more first learning data subsets. The processor 110 inputs at least a portion of the first training data set (for example, a training data set combining data sampled from each of the first training data subsets) to the first model to learn at least one training epoch can do it The processor 110 may input a training data set obtained by combining data sampled from each of the one or more first training data subsets to the first model to learn by at least one training epoch. For each training epoch, the processor 110 may obtain an updated weight of the first model. For example, the weight of the first model obtained by the processor 110 after 1 training epoch may be different from the weight of the first model obtained by the processor 110 after 2 training epochs. The foregoing is merely an example, and the present disclosure is not limited thereto.

According to an embodiment of the present disclosure, the processor 110 may perform a first learning operation 210 for learning the first model for each learning epoch with respect to the first model, respectively. As shown in FIG. 3 , the first learning operation 210 may include 4 learning epochs. The processor 110 may acquire the first model 1 221 after the training epoch 1 211 . The processor 110 may acquire the first model 2 223 after the training epoch 2 213 . After the first training epoch, the weights included in the model are updated, so the weights included in the first model 1 221 and the first model 2 223 may be different from each other. According to an embodiment of the present disclosure, the processor 110 may perform a second learning operation after a first learning operation of performing one epoch learning. According to another embodiment of the present disclosure, the processor 110 may perform a second learning operation after a first learning operation of performing 2 epoch learning. According to another embodiment of the present disclosure, the processor 110 may perform a second learning operation after a first learning operation of performing 3 epoch learning. Depending on how many learning epochs the first learning operation is performed, the weights of the first model that have undergone the first learning operation may be different. The processor 110 may perform a second learning operation for each of the first models learned with different epochs in the first learning operation. Through this, the processor 110 may determine the learning epoch of the first learning operation in which the loss value of the first model that has undergone the second learning operation is minimized. The foregoing is merely an example, and the present disclosure is not limited thereto.

Hereinafter, a detailed description of the second learning operation among the operations of pre-learning the first model will be described in detail with reference to FIG. 3 .

According to an embodiment of the present disclosure, the processor 110 may perform a second learning operation 230 for learning each of the first models that have undergone one or more first learning operations having different learning epochs. The second learning operation 230 may include an operation of inputting each of the first training data subsets corresponding to different tasks to the first model that has undergone the first learning operation 210 to learn by a predetermined second learning epoch. can The second learning operation may be an operation for determining a pre-trained first model. This is because the predetermined first learning epoch is determined based on the loss value of the first model that has undergone the second learning operation. For example, the loss value of the first model having a first predetermined training epoch of 1 and undergoing the second training operation is 0.7, and the loss value of the first model having a predetermined first training epoch of 2 and undergoing the second training operation is 0.6, the first predetermined training epoch is 3, the loss value of the first model that has undergone the second training operation is 0.5, the predetermined first training epoch is 4, and the loss value of the first model that has undergone the second training operation is 0.7 cases may exist. In this case, the processor 110 may determine a predetermined first learning epoch in which the loss value of the first model that has undergone the second learning operation is minimized as 3 . Accordingly, the pre-trained first model may be the first model in which learning is performed by the training epoch that minimizes the loss value of the first model that has undergone the second learning process. The foregoing is merely an example, and the present disclosure is not limited thereto.

According to an embodiment of the present disclosure, the training epoch inputs the training input data to the input node for all training data included in the training data set, respectively, compares the labeled data with the output of the model to derive an error, By propagating the derived error from an output layer of one or more network functions of the model to an input layer through one or more hidden layers, it may be an operation of updating a weight set in each link. The foregoing is merely an example, and the present disclosure is not limited thereto.

According to an embodiment of the present disclosure, the second predetermined learning epoch may include an epoch determined by an operator and/or an epoch automatically determined by the computing device. The second predetermined learning epoch may be the same as or different from the first predetermined learning epoch. The second predetermined learning epoch may include 1 epoch, 2 epochs, 3 epochs, and the like. The foregoing is merely an example, and the present disclosure is not limited thereto.

According to an embodiment of the present disclosure, the second learning operation 230 may include an operation of learning the model by inputting the first training data subset to the first model, respectively. The second learning operation 230 may include an operation for learning the model by inputting the first training data subset for each task to the first model that has undergone the first learning operation 210 . Through this, the processor 110 may acquire one or more models learned for each task through a second learning operation 230 with respect to the first model that has undergone the first learning operation. For example, the processor 110 may acquire the first model 4 227 that has undergone the first learning operation 210 by performing learning in 4 learning epochs on the first model. Next, the processor 110 generates the training data subset 1 corresponding to the task 1 250 , the training data subset 2 corresponding to the task 2 270 , and the training data subset 3 corresponding to the task 3 290 . The model may be trained by inputting the first model 4 227 that has undergone the first learning operation 210, respectively. In this case, the processor 110 may acquire three learned first models each having different weights for each task. The foregoing is merely an example, and the present disclosure is not limited thereto.

According to an embodiment of the present disclosure, the loss value of the one or more first models that have undergone the second learning operation may include task performance accuracy of each of the one or more first models that have undergone the second learning operation. The loss values of the one or more first models that have undergone the second learning operation may include respective loss values of the first models that have undergone the second learning operation corresponding to one or more tasks. The processor 110 may obtain a loss value of the learned first model having different weights, respectively. For example, as shown in FIG. 3 , the processor 110 has a loss value of 1 (257) 0.2 for task 1 (250), a loss value of 2 (277) 0.3 for task 2 (270), and task 3 (290). ), a loss value of 3 (297) 0.25 can be obtained. The processor 110 may acquire the pre-trained first model having the smallest loss value for all tasks through the first learning operation and the second learning operation. That is, it is possible to obtain the learning epoch in the first learning operation having the smallest loss value for all tasks. For example, when the predetermined first training epoch is 5, the loss value 1 (257) 0.1 for the task 1 (250), the loss value 2 (277) 0.2 for the task 2 (270), and the task 3 (290) ), a loss value of 3 (297) of 0.15 can be obtained. In this case, the processor 110 may determine the first model trained by 5 training epochs as the pre-trained first model. Through this, the processor 110 may obtain a pre-trained first model having excellent performance for all tasks included in the first training data set. The foregoing is merely an example, and the present disclosure is not limited thereto.

According to an embodiment of the present disclosure, the processor 110 may pre-train the first model. The processor 110 may determine the end of learning of the first learning operation based on the loss values of one or more first models that have undergone the second learning operation. The learning end of the first learning operation may be determined based on a predetermined first learning epoch. The predetermined first training epoch may be determined based on a loss value of the first model that has undergone the second training operation. The loss value of the first model that has undergone the second training operation may be different for each task included in the first training data set. Accordingly, the loss value of the first model that has undergone the second learning operation may include a statistical value of the loss value for each task included in the first training data set. The statistical value of the loss value for each task may include, for example, a simple summation, a weighted summation, a standard deviation, a statistical value considering variance, a median value, an average value, and the like. For example, as shown in FIG. 3 , the processor 110 has a loss value 1 257 corresponding to task 1 250 , a loss value 2 277 corresponding to task 2 270 , and task 3 290 . ), it is possible to determine the end of the learning of the first learning operation based on the statistical value of the loss value 3 (297) corresponding to the. That is, when the statistical value is equal to or less than a predetermined criterion, the processor 110 may determine the end of learning of the first learning operation. Accordingly, the processor 110 may determine a learning epoch for obtaining the pre-trained first model based on the loss value. The foregoing is merely an example, and the present disclosure is not limited thereto.

According to an embodiment of the present disclosure, the processor 110 may determine the first model learned by the determined learning epoch as the pre-trained first model. That is, the processor 110 may determine the end of learning of the first learning operation, and at the same time may also determine a learning epoch. For example, the learning end of the first learning operation may be 4 learning epochs. As shown in FIG. 3 , the learning epoch 4 217 may be the determined learning end point. In this case, the model having the weight included in the first model 4 227 may be the pre-trained first model. Accordingly, the processor 110 may determine the model at the time of the learning end (ie, the learning epoch at which the learning ends) of the first learning operation 210 as the pre-trained first model. The foregoing is merely an example, and the present disclosure is not limited thereto.

Hereinafter, a process of learning the second model will be described in detail with reference to FIG. 4 .

According to an embodiment of the present disclosure, the processor 110 may train a second model including at least a portion of the pre-trained first model. The processor 110 may train the model by inputting the second training data set to the second model. The second model may include a machine learning model including at least a portion of the pre-trained first model. The second model may include a deep learning model including at least a portion of the pre-trained first model. The trained second model may include a model obtained by transfer learning of the pre-trained first model. Accordingly, the processor 110 may acquire the second model capable of performing a specific task even using a training data set that is insufficient to train the model. The second model may include at least one of at least a portion of the pre-trained first model as a model for performing a single task or a regularization layer specialized for a single task. In another embodiment, the second model may be a model including layers other than the regularization layer included in the pre-trained first model. For example, the second model may include a regularization layer specialized for a single task instead of the regularization layer included in the pre-trained first model. The foregoing is merely an example, and the present disclosure is not limited thereto.

According to an embodiment of the present disclosure, the second training data set may include a training data set in which the second training data set includes a smaller number of data than the first training data set. That is, the second training data set may be a training data set having an insufficient number of data to train the model. For example, when the processor 110 trains the model using only the second training data set, the prediction accuracy of the model may be low. Accordingly, the processor 110 may train the model by inputting the second training data set to the pre-trained first model as in the present disclosure. Since the model trained by inputting the second training data set to the pre-trained first model has information about more tasks, it may be helpful for the processor 110 to perform a single task using the model. Through this, the processor 110 may obtain a model with high prediction accuracy for a single task. According to an embodiment of the present disclosure, the second training data set may include a training data set corresponding to a single task. The single task may be one of the tasks included in the first training data set. Also, a single task may be a new task not included in the first training data set. The foregoing is merely an example, and the present disclosure is not limited thereto.

According to an embodiment of the present disclosure, the processor 110 may train a second model including at least a portion of the pre-trained first model. The second model is a model for performing a single task included in the second training data set, and may include at least one of at least a portion of the pre-trained first model or at least one of a regularization layer specialized for a single task. When the processor 110 obtains the predicted value using the second model including at least a portion of the first model, an overfitting phenomenon may occur. Overfitting may mean a phenomenon in which errors on actual data increase due to over-training of a part of the training data. According to the present disclosure, when using the pre-trained first model and a regularization layer specialized for a single task, the processor 110 may obtain a second model with reduced errors for real data. According to an embodiment of the present disclosure, the normalization layer may include a layer for generating a normalized output by receiving an output of a previous layer. The normalization layer may be a layer used to mitigate overfitting. The normalization layer may include a general normalization layer and/or a batch normalization layer. The general normalization layer may be a layer that generates a normalized output based on a mean, variance, or standard deviation for each data. The batch normalization layer may be a layer that generates a normalized output based on an average, variance, or standard deviation for each batch including a plurality of data. The foregoing is merely an example, and the present disclosure is not limited thereto.

According to an embodiment of the present disclosure, as shown in FIG. 4 , the second model 1 310 may include at least a portion 311 of the pre-trained first model and a fourth task normalization layer 313 . . The fourth task may be a new task other than the task included in the first training data set. The fourth task normalization layer 313 may be a normalization layer specialized for the fourth task. Also, the fourth task normalization layer 313 may be a batch normalization layer specialized for the fourth task. The fourth task normalization layer 313 may be disposed behind the pre-trained first model. Also, the fourth task normalization layer 313 may be disposed between layers included in the pre-trained first model. For example, a layer included in the pre-trained first model may be a convolutional layer or a fully connected layer. In this case, the fourth task normalization layer may be disposed after the convolutional layer and before the fully connected layer. The second training data set 315 corresponding to the fourth task may have an insufficient number of training data to train the model. However, the processor 110 may obtain a second model having high prediction accuracy with respect to the fourth task by using at least a portion of the pre-trained first model and/or a regularization layer specialized for a single task. The foregoing is merely an example, and the present disclosure is not limited thereto.

According to an embodiment of the present disclosure, as shown in FIG. 4 , the second model 2 330 may include at least a portion 331 of the pre-trained first model and a fifth task normalization layer 333 . . The fifth task may be a task included in the first training data set. According to an embodiment of the present disclosure, the fifth task normalization layer 333 may be a batch normalization layer specialized for the fifth task. The fifth task normalization layer 333 may be disposed behind the pre-trained first model. Also, the fifth task normalization layer 333 may be disposed between layers included in the pre-trained first model. The foregoing is merely an example, and the present disclosure is not limited thereto.

According to an embodiment of the present disclosure, the processor 110 may train a second model including at least a portion of the pre-trained first model. The processor 110 may update the weight included in at least a part of the pre-trained first model and/or the weight included in the normalization layer in the process of training the second model using the second training data set. In the process of learning the second model 1 310 using the second training data set 315 corresponding to the fourth task, the processor 110 includes a weight included in at least a portion 311 of the first model pre-trained. and/or the weight included in the fourth task normalization layer 313 may be updated. The foregoing is merely an example, and the present disclosure is not limited thereto.

According to an embodiment of the present disclosure, the processor 110 may train a second model including at least a portion of the pre-trained first model. The processor 110 may fix a weight corresponding to at least a part of the pre-trained first model and train only a regularization layer specialized for a single task. The processor 110 may not update the weights included in at least some of the pre-trained first models 311 in the process of learning the second model 1 310 . Accordingly, the amount of computation to be performed by the processor 110 may be reduced and the learning time of the model may be shortened. The processor 110 may update only the weights included in the fourth task normalization layer 313 according to the learning epoch in a situation in which at least a portion 311 of the pre-trained first model is not updated. The foregoing is merely an example, and the present disclosure is not limited thereto.

According to the present disclosure, the processor 110 may perform a new task by using the learned model in a situation in which the number of training data for the new task is insufficient. More specifically, the processor 110 may acquire a model with high prediction accuracy for a new task by acquiring the pre-trained first model and then transferring the second model to fit the new task. In addition, according to the present disclosure, by including a regularization layer specialized for a single task in the second model and learning, it is possible to alleviate the overfitting phenomenon of the learned model. By using a regularization layer specialized for a single task, it is possible to obtain the output value of the model in which the overfitting phenomenon is alleviated even for a new task that is similar to the previously learned task. Through this, the processor 110 may obtain a model with high prediction accuracy even if the number of training data for a new task is insufficient.

2 is a schematic diagram exemplarily illustrating a neural network in a method for training a neural network, according to an embodiment of the present disclosure.

Throughout this specification, computational model, neural network, network function, and neural network may be used interchangeably. A neural network may be composed of a set of interconnected computational units, which may generally be referred to as nodes. These nodes may also be referred to as neurons. A neural network is configured by including at least one or more nodes. Nodes (or neurons) constituting the neural networks may be interconnected by one or more links.

In the neural network, one or more nodes connected through a link may relatively form a relationship between an input node and an output node. The concepts of an input node and an output node are relative, and any node in an output node relationship with respect to one node may be in an input node relationship in a relationship with another node, and vice versa. As described above, an input node-to-output node relationship may be created around a link. One or more output nodes may be connected to one input node through a link, and vice versa.

In the relationship between the input node and the output node connected through one link, the value of the output node may be determined based on data input to the input node. Here, a node interconnecting the input node and the output node may have a weight. The weight may be variable, and may be changed by the user or algorithm in order for the neural network to perform a desired function. For example, when one or more input nodes are interconnected to one output node by respective links, the output node sets values input to input nodes connected to the output node and links corresponding to the respective input nodes. An output node value may be determined based on the weight.

As described above, in a neural network, one or more nodes are interconnected through one or more links to form an input node and an output node relationship within the neural network. The characteristics of the neural network may be determined according to the number of nodes and links in the neural network, the correlation between the nodes and the links, and the value of a weight assigned to each of the links. For example, when the same number of nodes and links exist and there are two neural networks having different weight values between the links, the two neural networks may be recognized as different from each other.

A neural network may include one or more nodes. Some of the nodes constituting the neural network may configure one layer based on distances from the initial input node. For example, a set of nodes having a distance n from the initial input node may constitute n layers. The distance from the initial input node may be defined by the minimum number of links that must be traversed to reach the corresponding node from the initial input node. However, the definition of such a layer is arbitrary for description, and the order of the layer in the neural network may be defined in a different way from the above. For example, a layer of nodes may be defined by a distance from the final output node.

The initial input node may refer to one or more nodes to which data is directly input without going through a link in a relationship with other nodes among nodes in the neural network. Alternatively, in a relationship between nodes based on a link in a neural network, it may mean nodes that other input nodes connected by a link do not have. Similarly, the final output node may refer to one or more nodes that do not have an output node in relation to other nodes among nodes in the neural network. In addition, the hidden node may mean nodes constituting the neural network other than the first input node and the last output node. The neural network according to an embodiment of the present disclosure may be a neural network in which the number of nodes in the input layer may be the same as the number of nodes in the output layer, and the number of nodes decreases and then increases again as progresses from the input layer to the hidden layer. can In addition, the neural network according to another embodiment of the present disclosure may be a neural network in which the number of nodes in the input layer may be less than the number of nodes in the output layer, and the number of nodes decreases as the number of nodes progresses from the input layer to the hidden layer. have. In addition, the neural network according to another embodiment of the present disclosure may be a neural network in which the number of nodes in the input layer may be greater than the number of nodes in the output layer, and the number of nodes increases as the input layer progresses to the hidden layer. can The neural network according to another embodiment of the present disclosure may be a neural network in a combined form of the aforementioned neural networks.

A deep neural network (DNN) may refer to a neural network including a plurality of hidden layers in addition to an input layer and an output layer. Deep neural networks can be used to identify the latent structures of data. In other words, it can identify the potential structure of photos, texts, videos, voices, and music (for example, what objects are in the photos, what the text and emotions are, what the texts and emotions are, etc.) . Deep neural networks include convolutional neural networks (CNNs), recurrent neural networks (RNNs), auto encoders, generative adversarial networks (GANs), and restricted Boltzmann machines (RBMs). boltzmann machine), a deep belief network (DBN), a Q network, a U network, a Siamese network, and the like. The description of the deep neural network described above is only an example, and the present disclosure is not limited thereto.

In an embodiment of the present disclosure, the network function may include an auto-encoder. The auto-encoder may be a kind of artificial neural network for outputting output data similar to input data. The auto encoder may include at least one hidden layer, and an odd number of hidden layers may be disposed between the input/output layers. The number of nodes in each layer may be reduced from the number of nodes of the input layer to an intermediate layer called the bottleneck layer (encoding), and then expanded symmetrically with reduction from the bottleneck layer to the output layer (symmetrically with the input layer). In this case, although the dimensionality reduction layer and the dimension reconstruction layer are illustrated as being symmetrical in the example of FIG. 2 , the present disclosure is not limited thereto, and nodes of the dimension reduction layer and the dimension reconstruction layer may or may not be symmetrical. The auto-encoder can perform non-linear dimensionality reduction. The number of input layers and output layers may correspond to the number of sensors remaining after pre-processing of input data. In the auto-encoder structure, the number of nodes of the hidden layer included in the encoder may have a structure that decreases as the distance from the input layer increases. If the number of nodes in the bottleneck layer (the layer with the fewest nodes between the encoder and the decoder) is too small, a sufficient amount of information may not be conveyed, so a certain number or more (e.g., more than half of the input layer, etc.) ) may be maintained.

The neural network may be trained by at least one of supervised learning, unsupervised learning, and semi-supervised learning. The training of the neural network is to minimize the error in the output. In the training of a neural network, iteratively inputs the training data to the neural network, calculates the output and target errors of the neural network for the training data, and reduces the error of the neural network from the output layer of the neural network to the input layer. It is a process of updating the weight of each node in the neural network by backpropagation in the direction. In the case of teacher learning, learning data in which the correct answer is labeled in each learning data is used (ie, labeled learning data), and in the case of comparative learning, the correct answer may not be labeled in each learning data. That is, for example, the learning data in the case of teacher learning regarding data classification may be data in which categories are labeled in each of the learning data. Labeled training data is input to the neural network, and an error can be calculated by comparing the output (category) of the neural network with the label of the training data. As another example, in the case of comparison learning about data classification, an error may be calculated by comparing the input training data with the neural network output. The calculated error is back propagated in the reverse direction (ie, from the output layer to the input layer) in the neural network, and the connection weight of each node of each layer of the neural network may be updated according to the back propagation. A change amount of the connection weight of each node to be updated may be determined according to a learning rate. The computation of the neural network on the input data and the backpropagation of errors can constitute a learning cycle (epoch). The learning rate may be applied differently according to the number of repetitions of the learning cycle of the neural network. For example, in the early stage of training of a neural network, a high learning rate can be used to enable the neural network to quickly acquire a certain level of performance, thereby increasing efficiency, and using a low learning rate at the end of learning can increase accuracy.

In the training of neural networks, in general, the training data may be a subset of real data (that is, data to be processed using the trained neural network), and thus the error on the training data is reduced, but the error on the real data is reduced. There may be increasing learning cycles. Overfitting is a phenomenon in which errors on actual data increase by over-learning on training data as described above. For example, a phenomenon in which a neural network that has learned a cat by showing a yellow cat does not recognize that it is a cat when it sees a cat other than yellow may be a type of overfitting. Overfitting can act as a cause of increasing errors in machine learning algorithms. In order to prevent such overfitting, various optimization methods can be used. In order to prevent overfitting, methods such as increasing training data, regularization, or dropout in which a part of nodes in the network are omitted in the process of learning, may be applied.

3 is an exemplary diagram for explaining a process for pre-training the first model.

3 shows a first learning operation 210 and a second learning operation 230 . Also, the learning epoch proceeding in the first learning operation 210 and the learning epoch proceeding in the second learning operation 230 are exemplarily shown.

According to an embodiment of the present disclosure, the computing device 100 may perform a first learning operation 210 for learning the first model for each training epoch with respect to the first model, respectively. The first learning operation 210 may include 4 learning epochs. The computing device 100 may acquire the first model 1 221 after the training epoch 1 211 . The computing device 100 may acquire the first model 2 223 after the training epoch 2 213 . After the first training epoch, the weights included in the model are updated, so the weights included in the first model 1 221 and the first model 2 223 may be different from each other. According to an embodiment of the present disclosure, the computing device 100 may perform a second learning operation after a first learning operation of performing one epoch learning. According to another embodiment of the present disclosure, the computing device 100 may perform a second learning operation after a first learning operation of performing 2 epoch learning. According to another embodiment of the present disclosure, the computing device 100 may perform a second learning operation after a first learning operation of performing 3 epoch learning. Depending on how many learning epochs the first learning operation is performed, the weights of the first model that have undergone the first learning operation may be different. The computing device 100 may perform a second learning operation for each of the first models learned with different epochs in the first learning operation. Through this, the computing device 100 may determine the learning epoch of the first learning operation in which the loss value of the first model that has undergone the second learning operation is minimized. The foregoing is merely an example, and the present disclosure is not limited thereto.

According to an embodiment of the present disclosure, the second learning operation 230 may include an operation of learning the model by inputting the first training data subset to the first model, respectively. The second learning operation 230 may include an operation for learning the model by inputting the first training data subset for each task to the first model that has undergone the first learning operation 210 . Through this, the computing device 100 may acquire one or more models learned for each task through the second learning operation 230 with respect to the first model that has undergone the first learning operation. For example, the computing device 100 may acquire the first model 4 227 that has undergone the first learning operation 210 by performing learning in 4 training epochs on the first model. Next, the computing device 100 sets the training data subset 1 corresponding to the task 1 250 , the training data subset 2 corresponding to the task 2 270 , and the training data subset 3 corresponding to the task 3 290 . may be input to the first model 4 227 that has undergone the first learning operation 210, respectively, to train the model. In this case, the computing device 100 may acquire the learned first model having different weights for each of the three tasks. The foregoing is merely an example, and the present disclosure is not limited thereto.

According to an embodiment of the present disclosure, the loss value of the one or more first models that have undergone the second learning operation may include task performance accuracy of each of the one or more first models that have undergone the second learning operation. The loss values of the one or more first models that have undergone the second learning operation may include respective loss values of the first models that have undergone the second learning operation corresponding to one or more tasks. The processor 110 may obtain a loss value of the learned first model having different weights, respectively. For example, as shown in FIG. 3 , the computing device 100 has a loss value of 1 ( 257 ) 0.2 for task 1 ( 250 ), a loss value of 2 ( 277 ) 0.3 for task 2 ( 270 ), and task 3 ( For 290), a loss value of 3 (297) 0.25 can be obtained. The computing device 100 may acquire the pre-trained first model having the smallest loss value for all tasks through the first learning operation and the second learning operation. That is, it is possible to obtain a learning epoch in the first learning operation to have the smallest loss value for all tasks. For example, when the predetermined first training epoch is 5, the loss value 1 (257) 0.1 for the task 1 (250), the loss value 2 (277) 0.2 for the task 2 (270), and the task 3 (290) ), a loss value of 3 (297) of 0.15 can be obtained. In this case, the computing device 100 may determine the first model trained by 5 training epochs as the pre-trained first model. Through this, the computing device 100 may obtain a pre-trained first model having excellent performance for all tasks included in the first training data set. The foregoing is merely an example, and the present disclosure is not limited thereto.

4 is an exemplary diagram for explaining an operation for learning a second model.

According to an embodiment of the present disclosure, as shown in FIG. 4 , the second model 1 310 may include at least a portion 311 of the pre-trained first model and a fourth task normalization layer 313 . . The fourth task may be a new task other than the task included in the first training data set. The fourth task normalization layer 313 may be a normalization layer specialized for the fourth task. Also, the fourth task normalization layer 313 may be a batch normalization layer specialized for the fourth task. The fourth task normalization layer 313 may be disposed behind the pre-trained first model. Also, the fourth task normalization layer 313 may be disposed between layers included in the pre-trained first model. For example, a layer included in the pre-trained first model may be a convolutional layer or a fully connected layer. In this case, the fourth task normalization layer may be disposed after the convolutional layer and before the fully connected layer. The second training data set 315 corresponding to the fourth task may have an insufficient number of training data to train the model. However, the computing device 100 may obtain a second model having high prediction accuracy with respect to the fourth task by using at least a portion of the pre-trained first model and/or a regularization layer specialized for a single task. The foregoing is merely an example, and the present disclosure is not limited thereto.

According to an embodiment of the present disclosure, as shown in FIG. 4 , the second model 2 330 may include at least a portion 331 of the pre-trained first model and a fifth task normalization layer 333 . . The fifth task may be a task included in the first training data set. According to an embodiment of the present disclosure, the fifth task normalization layer 333 may be a batch normalization layer specialized for the fifth task. The fifth task normalization layer 333 may be disposed behind the pre-trained first model. Also, the fifth task normalization layer 333 may be disposed between layers included in the pre-trained first model. The foregoing is merely an example, and the present disclosure is not limited thereto.

According to an embodiment of the present disclosure, the computing device 100 may train a second model including at least a portion of the pre-trained first model. The computing device 100 may update the weight included in at least a part of the pre-trained first model and/or the weight included in the regularization layer in a process of learning the second model using the second training data set. The computing device 100 includes at least a portion 311 of the first model pre-trained in the process of learning the second model 1 310 using the second training data set 315 corresponding to the fourth task. Weights and/or weights included in the fourth task normalization layer 313 may be updated. The foregoing is merely an example, and the present disclosure is not limited thereto.

According to an embodiment of the present disclosure, the computing device 100 may train a second model including at least a portion of the pre-trained first model. The computing device 100 may fix a weight corresponding to at least a part of the pre-trained first model and train only a regularization layer specialized for a single task. The computing device 100 may not update the weights included in at least some of the pre-trained first models 311 in the process of learning the second model 1 310 . Accordingly, the amount of computation to be performed by the computing device 100 may be reduced and the training time of the model may be shortened. The computing device 100 may update only the weights included in the fourth task normalization layer 313 according to the learning epoch in a situation in which at least a portion 311 of the pre-trained first model is not updated. The foregoing is merely an example, and the present disclosure is not limited thereto.

5 is a flowchart for training a neural network.

According to an embodiment of the present disclosure, the computing device 100 may input ( 410 ) a first training data set including one or more tasks to the first model.

According to an embodiment of the present disclosure, the first training data set may include one or more first training data subsets for each task.

According to an embodiment of the present disclosure, the task may include at least one of an image data processing task, a voice data processing task, a natural language processing task, and a time series data processing task as a processing operation using the learned model.

According to an embodiment of the present disclosure, the image data processing task includes a task of classifying input image data based on a predetermined criterion, a task of segmenting input image data, and a task of detecting anomalies in the input image data. Alternatively, it may include at least one of a task of recognizing an object from input image data.

According to an embodiment of the present disclosure, the computing device 100 may pre-train the first model ( 420 ).

According to an embodiment of the present disclosure, the loss value of the one or more first models that have undergone the second learning operation is the task performance accuracy of each of the one or more first models that have undergone the second learning operation, and a second corresponding to the one or more tasks. Each of the loss values of the first model that has undergone the learning operation may be included.

According to an embodiment of the present disclosure, the computing device 100 may train a second model including at least a portion of the pre-trained first model ( 430 ).

According to an embodiment of the present disclosure, the second training data set is a training data set including a smaller number of data than the first training data set, and may include a training data set corresponding to a single task.

According to an embodiment of the present disclosure, the second model is a model for performing a single task included in the second training data set, and at least one of at least a portion of the pre-trained first model or at least one of a regularization layer specialized for a single task may include.

6 is a block diagram illustrating a module for learning a neural network.

A method for learning a neural network according to an embodiment of the present disclosure may be implemented by the following modules. a module 510 for inputting a first training data set comprising one or more tasks into a first model; a module 520 for pre-training the first model; and a module 530 for training a second model including at least a portion of the pre-trained first model.

In an alternative embodiment for training a neural network, the module 520 for pre-training the first model is configured to: minimize the loss of the first model obtained by inputting the first training data set to the first model. 1 It may include a module for pre-training the model.

In an alternative embodiment, the module 520 for pre-training the first model includes: a first training module for respectively training the first model for each training epoch with respect to the first model; and a second learning module for learning each of the first models that have undergone one or more first learning operations having different learning epochs.

In an alternative embodiment for training the neural network, the first training module may include a module for inputting at least a portion of the first training data set to the first model to train by a predetermined first training epoch.

In an alternative embodiment for training a neural network, the second training module is configured to input each of the first training data subsets corresponding to different tasks to the first model subjected to the first training operation to learn by a second predetermined training epoch. It may include a module for doing so.

In an alternative embodiment for training a neural network, the module 520 for pre-training the first model determines the end of learning of the first learning operation based on loss values of one or more first models that have undergone the second learning operation. module for; and a module for determining the first model trained by the determined learning epoch as the pre-trained first model.

In an alternative embodiment for training a neural network, module 530 for training a second model comprising at least a portion of a pre-trained first model is configured to input a second training data set to the second model for training. It can contain modules.

In an alternative embodiment for training a neural network, the module 530 for training a second model comprising at least a portion of the pre-trained first model is configured to calculate weights corresponding to at least a portion of the pre-trained first model. It can be fixed and include a module for training only a single-task-specific regularization layer.

According to an embodiment of the present disclosure, a module for learning a neural network may be implemented by means, circuitry or logic for implementing a computing device. Those skilled in the art will further appreciate that the various illustrative logical blocks, configurations, modules, circuits, means, logics, and algorithm steps described in connection with the embodiments disclosed herein may be combined with electronic hardware, computer software, or combinations of both. It should be recognized that it can be implemented as To clearly illustrate this interchangeability of hardware and software, various illustrative components, blocks, configurations, means, logics, modules, circuits, and steps have been described above generally in terms of their functionality. Whether such functionality is implemented as hardware or software depends upon the particular application and design constraints imposed on the overall system. Skilled artisans may implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present disclosure.

7 is a simplified, general schematic diagram of an exemplary computing environment in which embodiments of the present disclosure may be implemented.

Although the present disclosure has been described above generally as being capable of being implemented by a computing device, those skilled in the art will appreciate that the present disclosure is a combination of hardware and software and/or in combination with computer-executable instructions and/or other program modules that may be executed on one or more computers It will be appreciated that it can be implemented as a combination.

Generally, program modules include routines, programs, components, data structures, etc. that perform particular tasks or implement particular abstract data types. In addition, those skilled in the art will appreciate that the methods of the present disclosure are suitable for single-processor or multiprocessor computer systems, minicomputers, mainframe computers as well as personal computers, handheld computing devices, microprocessor-based or programmable consumer electronics, etc. (each of which It will be appreciated that other computer system configurations may be implemented, including those that may operate in connection with one or more associated devices.

The described embodiments of the present disclosure may also be practiced in distributed computing environments where certain tasks are performed by remote processing devices that are linked through a communications network. In a distributed computing environment, program modules may be located in both local and remote memory storage devices.

Computers typically include a variety of computer-readable media. Any medium accessible by a computer can be a computer readable medium, and such computer readable media includes volatile and nonvolatile media, transitory and non-transitory media, removable and non-transitory media. including removable media. By way of example, and not limitation, computer-readable media may include computer-readable storage media and computer-readable transmission media. Computer-readable storage media includes volatile and non-volatile media, transitory and non-transitory media, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. includes media. A computer-readable storage medium may be RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital video disk (DVD) or other optical disk storage device, magnetic cassette, magnetic tape, magnetic disk storage device, or other magnetic storage device. device, or any other medium that can be accessed by a computer and used to store the desired information.

Computer readable transmission media typically embodies computer readable instructions, data structures, program modules or other data, etc. in a modulated data signal such as a carrier wave or other transport mechanism, and Includes any information delivery medium. The term modulated data signal means a signal in which one or more of the characteristics of the signal is set or changed so as to encode information in the signal. By way of example, and not limitation, computer-readable transmission media includes wired media such as a wired network or direct-wired connection, and wireless media such as acoustic, RF, infrared, and other wireless media. Combinations of any of the above are also intended to be included within the scope of computer-readable transmission media.

An example environment 1100 implementing various aspects of the disclosure is shown including a computer 1102 , the computer 1102 including a processing unit 1104 , a system memory 1106 , and a system bus 1108 . do. The system bus 1108 couples system components, including but not limited to system memory 1106 , to the processing device 1104 . The processing device 1104 may be any of a variety of commercially available processors. Dual processor and other multiprocessor architectures may also be used as processing unit 1104 .

The system bus 1108 may be any of several types of bus structures that may further be interconnected to a memory bus, a peripheral bus, and a local bus using any of a variety of commercial bus architectures. System memory 1106 includes read only memory (ROM) 1110 and random access memory (RAM) 1112 . A basic input/output system (BIOS) is stored in non-volatile memory 1110, such as ROM, EPROM, EEPROM, etc., which is the basic input/output system (BIOS) that helps transfer information between components within computer 1102, such as during startup contains routines. RAM 1112 may also include high-speed RAM, such as static RAM, for caching data.

The computer 1102 may also be configured for external use within an internal hard disk drive (HDD) 1114 (eg, EIDE, SATA) - this internal hard disk drive 1114 may also be configured for external use within a suitable chassis (not shown). Yes-, magnetic floppy disk drive (FDD) 1116 (eg, for reading from or writing to removable diskette 1118), and optical disk drive 1120 (eg, CD-ROM) for reading from, or writing to, disk 1122, or other high capacity optical media such as DVD. The hard disk drive 1114 , the magnetic disk drive 1116 , and the optical disk drive 1120 are connected to the system bus 1108 by the hard disk drive interface 1124 , the magnetic disk drive interface 1126 , and the optical drive interface 1128 , respectively. ) can be connected to The interface 1124 for implementing an external drive includes at least one or both of Universal Serial Bus (USB) and IEEE 1394 interface technologies.

These drives and their associated computer-readable media provide non-volatile storage of data, data structures, computer-executable instructions, and the like. In the case of computer 1102, drives and media correspond to storing any data in a suitable digital format. Although the description of computer readable media above refers to HDDs, removable magnetic disks, and removable optical media such as CDs or DVDs, those skilled in the art will use zip drives, magnetic cassettes, flash memory cards, cartridges, etc. It will be appreciated that other tangible computer-readable media such as etc. may also be used in the exemplary operating environment and any such media may include computer-executable instructions for performing the methods of the present disclosure.

A number of program modules may be stored in the drive and RAM 1112 , including an operating system 1130 , one or more application programs 1132 , other program modules 1134 , and program data 1136 . All or portions of the operating system, applications, modules, and/or data may also be cached in RAM 1112 . It will be appreciated that the present disclosure may be implemented in various commercially available operating systems or combinations of operating systems.

A user may enter commands and information into the computer 1102 via one or more wired/wireless input devices, for example, a pointing device such as a keyboard 1138 and a mouse 1140 . Other input devices (not shown) may include a microphone, IR remote control, joystick, game pad, stylus pen, touch screen, and the like. Although these and other input devices are often connected to the processing unit 1104 through an input device interface 1142 that is connected to the system bus 1108, parallel ports, IEEE 1394 serial ports, game ports, USB ports, IR interfaces, It may be connected by other interfaces, etc.

A monitor 1144 or other type of display device is also coupled to the system bus 1108 via an interface, such as a video adapter 1146 . In addition to the monitor 1144, the computer typically includes other peripheral output devices (not shown), such as speakers, printers, and the like.

Computer 1102 may operate in a networked environment using logical connections to one or more remote computers, such as remote computer(s) 1148 via wired and/or wireless communications. Remote computer(s) 1148 may be workstations, computing device computers, routers, personal computers, portable computers, microprocessor-based entertainment devices, peer devices, or other common network nodes, and are typically connected to computer 1102 . Although it includes many or all of the components described for it, only memory storage device 1150 is shown for simplicity. The logical connections shown include wired/wireless connections to a local area network (LAN) 1152 and/or a larger network, eg, a wide area network (WAN) 1154 . Such LAN and WAN networking environments are common in offices and companies, and facilitate enterprise-wide computer networks, such as intranets, all of which can be connected to a worldwide computer network, for example, the Internet.

When used in a LAN networking environment, the computer 1102 is connected to the local network 1152 through a wired and/or wireless communication network interface or adapter 1156 . Adapter 1156 may facilitate wired or wireless communication to LAN 1152 , which also includes a wireless access point installed therein for communicating with wireless adapter 1156 . When used in a WAN networking environment, the computer 1102 may include a modem 1158, be connected to a communication computing device on the WAN 1154, or establish communications over the WAN 1154, such as over the Internet. have other means. A modem 1158 , which may be internal or external and a wired or wireless device, is coupled to the system bus 1108 via a serial port interface 1142 . In a networked environment, program modules described for computer 1102 , or portions thereof, may be stored in remote memory/storage device 1150 . It will be appreciated that the network connections shown are exemplary and other means of establishing a communication link between the computers may be used.

Computer 1102 may be associated with any wireless device or object that is deployed and operates in wireless communication, for example, printers, scanners, desktop and/or portable computers, portable data assistants (PDAs), communications satellites, wireless detectable tags. It operates to communicate with any device or place, and phone. This includes at least Wi-Fi and Bluetooth wireless technologies. Accordingly, the communication may be a predefined structure as in a conventional network or may simply be an ad hoc communication between at least two devices.

Wi-Fi (Wireless Fidelity) makes it possible to connect to the Internet, etc. without a wire. Wi-Fi is a wireless technology such as cell phones that allows these devices, eg, computers, to transmit and receive data indoors and outdoors, ie anywhere within range of a base station. Wi-Fi networks use a radio technology called IEEE 802.11 (a, b, g, etc.) to provide secure, reliable, and high-speed wireless connections. Wi-Fi can be used to connect computers to each other, to the Internet, and to wired networks (using IEEE 802.3 or Ethernet). Wi-Fi networks may operate in unlicensed 2.4 and 5 GHz radio bands, for example at 11 Mbps (802.11a) or 54 Mbps (802.11b) data rates, or in products that include both bands (dual band). .

One of ordinary skill in the art of this disclosure will understand that information and signals may be represented using any of a variety of different technologies and techniques. For example, data, instructions, instructions, information, signals, bits, symbols and chips that may be referenced in the above description are voltages, currents, electromagnetic waves, magnetic fields or particles, optical fields particles or particles, or any combination thereof.

Those of ordinary skill in the art of the present disclosure will recognize that the various illustrative logical blocks, modules, processors, means, circuits, and algorithm steps described in connection with the embodiments disclosed herein include electronic hardware, (convenience For this purpose, it will be understood that it may be implemented by various forms of program or design code (referred to herein as software) or a combination of both. To clearly illustrate this interchangeability of hardware and software, various illustrative components, blocks, modules, circuits, and steps have been described above generally in terms of their functionality. Whether such functionality is implemented as hardware or software depends upon the particular application and design constraints imposed on the overall system. A person skilled in the art of the present disclosure may implement the described functionality in various ways for each specific application, but such implementation decisions should not be interpreted as a departure from the scope of the present disclosure.

The various embodiments presented herein may be implemented as methods, apparatus, or articles of manufacture using standard programming and/or engineering techniques. The term article of manufacture includes a computer program, carrier, or media accessible from any computer-readable storage device. For example, computer-readable storage media include magnetic storage devices (eg, hard disks, floppy disks, magnetic strips, etc.), optical disks (eg, CDs, DVDs, etc.), smart cards, and flash drives. memory devices (eg, EEPROMs, cards, sticks, key drives, etc.). Also, various storage media presented herein include one or more devices and/or other machine-readable media for storing information.

It is understood that the specific order or hierarchy of steps in the presented processes is an example of exemplary approaches. Based on design priorities, it is understood that the specific order or hierarchy of steps in the processes may be rearranged within the scope of the present disclosure. The appended method claims present elements of the various steps in a sample order, but are not meant to be limited to the specific order or hierarchy presented.

The description of the presented embodiments is provided to enable any person skilled in the art to make or use the present disclosure. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other embodiments without departing from the scope of the present disclosure. Thus, the present disclosure is not intended to be limited to the embodiments presented herein, but is to be construed in the widest scope consistent with the principles and novel features presented herein.

A computer-readable medium storing a data structure is disclosed according to an embodiment of the present disclosure.

The data structure may refer to the organization, management, and storage of data that enables efficient access and modification of data. A data structure may refer to an organization of data to solve a specific problem (eg, data retrieval, data storage, and data modification in the shortest time). A data structure may be defined as a physical or logical relationship between data elements designed to support a particular data processing function. The logical relationship between data elements may include a connection relationship between data elements that a user thinks. Physical relationships between data elements may include actual relationships between data elements physically stored on a computer-readable storage medium (eg, a hard disk). A data structure may specifically include a set of data, relationships between data, and functions or instructions applicable to data. Through an effectively designed data structure, a computing device can perform an operation while using the resource of the computing device to a minimum. Specifically, the computing device may increase the efficiency of operations, reads, insertions, deletions, comparisons, exchanges, and retrievals through effectively designed data structures.

A data structure may be classified into a linear data structure and a non-linear data structure according to the type of the data structure. The linear data structure may be a structure in which only one piece of data is connected after one piece of data. The linear data structure may include a list, a stack, a queue, and a deck. A list may mean a set of data in which an order exists internally. The list may include a linked list. The linked list may be a data structure in which data is linked in such a way that each data is linked in a line with a pointer. In a linked list, a pointer may contain information about a link to the next or previous data. A linked list may be expressed as a single linked list, a doubly linked list, or a circularly linked list according to a shape. A stack can be a data enumeration structure with limited access to data. A stack can be a linear data structure in which data can be processed (eg, inserted or deleted) at only one end of the data structure. The data stored in the stack may be a data structure LIFO-Last in First Out. A queue is a data listing structure that allows limited access to data. Unlike a stack, the queue may be a data structure that comes out later (FIFO-First in First Out) as data stored later. A deck can be a data structure that can process data at either end of the data structure.

The nonlinear data structure may be a structure in which a plurality of data is connected after one data. The nonlinear data structure may include a graph data structure. A graph data structure may be defined as a vertex and an edge, and the edge may include a line connecting two different vertices. A graph data structure may include a tree data structure. The tree data structure may be a data structure in which one path connects two different vertices among a plurality of vertices included in the tree. That is, it may be a data structure that does not form a loop in the graph data structure.

Throughout this specification, computational model, neural network, network function, and neural network may be used interchangeably. (Hereinafter, the neural network is unified and described.) The data structure may include a neural network. And the data structure including the neural network may be stored in a computer-readable medium. Data structures, including neural networks, may also include data input to the neural network, weights of the neural network, hyperparameters of the neural network, data obtained from the neural network, activation functions associated with each node or layer of the neural network, and loss functions for learning the neural network. have. A data structure comprising a neural network may include any of the components disclosed above. That is, the data structure including the neural network includes all or all of the data input to the neural network, the weights of the neural network, hyperparameters of the neural network, data obtained from the neural network, the activation function associated with each node or layer of the neural network, and the loss function for training the neural network may be configured including any combination of In addition to the above-described configurations, a data structure including a neural network may include any other information that determines a characteristic of a neural network. In addition, the data structure may include all types of data used or generated in the operation process of the neural network, and is not limited to the above. Computer-readable media may include computer-readable recording media and/or computer-readable transmission media. A neural network may be composed of a set of interconnected computational units, which may generally be referred to as nodes. These nodes may also be referred to as neurons. A neural network is configured by including at least one or more nodes.

The data structure may include data input to the neural network. A data structure including data input to the neural network may be stored in a computer-readable medium. The data input to the neural network may include learning data input in a neural network learning process and/or input data input to the neural network in which learning is completed. Data input to the neural network may include pre-processing data and/or pre-processing target data. The preprocessing may include a data processing process for inputting data into the neural network. Accordingly, the data structure may include data to be pre-processed and data generated by pre-processing. The above-described data structure is merely an example, and the present disclosure is not limited thereto.

The data structure may include data input to or output from the neural network. A data structure including data input to or output from the neural network may be stored in a computer-readable medium. The data structure stored in the computer-readable medium may include data input in an inference process of the neural network or output data output as a result of inference of the neural network. In addition, since the data structure may include data processed by a specific data processing method, it may include data before and after processing. Accordingly, the data structure may include data to be processed and data processed through a data processing method.

The data structure may include the weights of the neural network. (In this specification, a weight and a parameter may be used interchangeably.) And a data structure including a weight of a neural network may be stored in a computer-readable medium. The neural network may include a plurality of weights. The weight may be variable, and may be changed by the user or algorithm in order for the neural network to perform a desired function. For example, when one or more input nodes are interconnected to one output node by respective links, the output node sets values input to input nodes connected to the output node and links corresponding to the respective input nodes. An output node value may be determined based on the parameter. The above-described data structure is merely an example, and the present disclosure is not limited thereto.

By way of example and not limitation, the weight may include a weight variable in a neural network learning process and/or a weight in which neural network learning is completed. The variable weight in the neural network learning process may include a weight at the start of the learning cycle and/or a variable weight during the learning cycle. The weight for which neural network learning is completed may include a weight for which a learning cycle is completed. Accordingly, the data structure including the weight of the neural network may include a data structure including the weight variable in the neural network learning process and/or the weight in which the neural network learning is completed. Therefore, it is assumed that the above-described weights and/or combinations of weights are included in the data structure including the weights of the neural network. The above-described data structure is merely an example, and the present disclosure is not limited thereto.

The data structure including the weights of the neural network may be stored in a computer-readable storage medium (eg, memory, hard disk) after being serialized. Serialization can be the process of converting a data structure into a form that can be reconstructed and used later by storing it on the same or a different computing device. The computing device may serialize the data structure to send and receive data over the network. A data structure including weights of the serialized neural network may be reconstructed in the same computing device or in another computing device through deserialization. The data structure including the weights of the neural network is not limited to serialization. Furthermore, the data structure including the weights of the neural network is a data structure to increase the efficiency of computation while using the resources of the computing device to a minimum (e.g., B-Tree, Trie, m-way search tree, AVL tree, Red-Black Tree). The foregoing is merely an example, and the present disclosure is not limited thereto.

The data structure may include hyper-parameters of the neural network. In addition, the data structure including the hyperparameters of the neural network may be stored in a computer-readable medium. The hyperparameter may be a variable variable by a user. Hyperparameters are, for example, learning rate, cost function, number of iterations of the learning cycle, weight initialization (e.g., setting the range of weight values subject to weight initialization), Hidden Unit The number (eg, the number of hidden layers, the number of nodes of the hidden layer) may be included. The above-described data structure is merely an example, and the present disclosure is not limited thereto.

Claims (18)

A computer program stored in a computer readable storage medium, wherein the computer program, when executed on one or more processors, performs the following operations for training a neural network, the operations comprising:
inputting a first training data set comprising one or more tasks to a first model;
pre-training the first model; and
training a second model including at least a portion of the pre-trained first model;
comprising,
A computer program stored in a computer-readable storage medium.
The method of claim 1,
The first training data set is
at least one subset of the first training data per task;
A computer program stored in a computer-readable storage medium.
3. The method of claim 2,
The task is
As a task to be processed using the learned model,
comprising at least one of an image data processing task, a speech data processing task, a natural language processing task, or a time series data processing task,
A computer program stored in a computer-readable storage medium.
4. The method of claim 3,
The image data processing task is
a task of classifying the input image data based on a predetermined criterion;
Segmentation task of input image data,
A task that detects anomaly in input image data, or
A task for recognizing an object from input image data
comprising at least one of
A computer program stored in a computer-readable storage medium.
The method of claim 1,
The operation of pre-training the first model is,
training the first model to perform one or more tasks included in the first training data set using the first model;
comprising,
A computer program stored in a computer-readable storage medium.
The method of claim 1,
The operation of pre-training the first model is,
pre-training the first model in order to minimize a loss of the first model obtained by inputting the first training data set to the first model;
comprising,
A computer program stored in a computer-readable storage medium.
7. The method of claim 6,
The operation of pre-training the first model is,
a first learning operation of learning the first model for each learning epoch with respect to the first model; and
a second learning operation for learning each of the first models that have undergone one or more first learning operations having different learning epochs;
comprising,
A computer program stored in a computer-readable storage medium.
8. The method of claim 7,
The first learning operation is,
inputting at least a portion of the first training data set to a first model and learning by a predetermined first training epoch;
comprising,
A computer program stored in a computer-readable storage medium.
8. The method of claim 7,
The second learning operation is
an operation of inputting each of the first training data subsets corresponding to different tasks to the first model that has undergone the first learning operation, and learning by a predetermined second training epoch;
comprising,
A computer program stored in a computer-readable storage medium.
8. The method of claim 7,
The operation of pre-training the first model is,
determining an end of learning of the first learning operation based on the loss values of one or more first models that have undergone the second learning operation; and
determining a first model learned by the determined learning epoch as a pre-trained first model;
comprising,
A computer program stored in a computer-readable storage medium.
10. The method of claim 9,
The loss value of one or more first models that have undergone the second learning operation is,
As the task performance accuracy of each of the one or more first models that have undergone the second learning operation,
Including each loss value of the first model that has undergone a second learning operation corresponding to one or more tasks,
A computer program stored in a computer-readable storage medium.
The method of claim 1,
The operation of learning a second model including at least a part of the pre-trained first model comprises:
inputting a second training data set to a second model to learn;
comprising,
A computer program stored in a computer-readable storage medium.
13. The method of claim 12,
The second training data set is,
A training data set comprising a smaller number of data than the first training data set, the training data set comprising:
comprising a training data set corresponding to a single task,
A computer program stored in a computer-readable storage medium.
13. The method of claim 12,
The second model is
A model for performing a single task included in a second training data set, comprising:
at least part of the pre-trained first model or
A regularization layer specialized for a single task
comprising at least one of
A computer program stored in a computer-readable storage medium.
The method of claim 1,
The operation of learning a second model including at least a part of the pre-trained first model comprises:
an operation of fixing weights corresponding to at least a part of the pre-trained first model and training only a regularization layer specialized for a single task;
comprising,
A computer program stored in a computer-readable storage medium.
A computer-readable recording medium storing a data structure corresponding to a parameter of a neural network that is at least partially updated in a learning process, wherein the operation of the neural network is based at least in part on the parameter, the learning process comprising:
inputting a first training data set comprising one or more tasks to a first model;
pre-training the first model; and
training a second model including at least a portion of the pre-trained first model;
containing,
A computer-readable recording medium having a data structure stored thereon.
A method for training a neural network, comprising:
inputting a first training data set comprising one or more tasks to a first model;
pre-training the first model; and
training a second model including at least a portion of the pre-trained first model;
containing,
A method for training a neural network.
A computing device for training a neural network, comprising:
one or more processors; and
a memory storing instructions executable by the one or more processors;
including,
The one or more processors,
inputting a first training data set comprising one or more tasks to a first model;
pre-train the first model, and
training a second model including at least a portion of the pre-trained first model,
computing device.

Images