KR20210030063A - System and method for constructing a generative adversarial network model for image classification based on semi-supervised learning - Google Patents

Abstract

The present invention relates to a method for constructing an adversarial image generative model based on semi-supervised learning. The method for constructing an adversarial image generative model includes the steps of: (a) obtaining a first loss function according to semi-supervised learning for a discriminator and learning to optimize the first loss function; (b) obtaining a second loss function for minimizing an Earth Mover′s distance (EM distance) for the discriminator and learning to optimize the second loss function; (c) obtaining a third loss function according to an existing adversarial image generative model for a generator and learning to optimize the third loss function; (d) obtaining a fourth loss function for minimizing the EM distance for the generator and learning to optimize the fourth loss function; and (e) classifying a result of the discriminator using a classifier and providing an output value.

Description

System and method for constructing a generative adversarial network model for image classification based on semi-supervised learning}

The present invention relates to a system and method for constructing a hostile image generation model, and more specifically, by combining a hostile image generation model (Semi-supervised GAN;'SGAN') and a Wassersteing GAN (WGAN) model based on semi-supervised learning. The present invention relates to a method of constructing an improved hostile image generation model for classification of composed images.

In general, classification algorithms based on deep learning train a model using a certain amount of training data, and classify what kind of image the test data is based on. Machine learning is divided into three types: supervised learning, unsupervised learning, and semi-supervised learning. Supervised learning is machine learning that infers a function from training data, and training data consists of an input value and a target value or a correct answer for it. Unsupervised learning belongs to the category of problems to find out how the data is structured, and does not include the target value or the correct answer for the input value of the training data. On the other hand, semi-supervised learning is a category of machine learning, in which both data with target values and undisplayed data are used for training.

Of these, supervised learning can expect better results with a sufficiently large amount of labeled data. However, according to the recent rapid development of information technology, a huge amount of data that cannot be compared to the past is accumulated, but among them, the labeled data for supervised learning is bound to occupy a very small proportion. This is because adding a label requires human labor or additional resources.

One of the ways to solve this problem is unsupervised learning. Unsupervised learning has the advantage of being able to use a myriad of data because it uses unlabeled data. However, there is a disadvantage in that the learning method is not clearly presented, the types are diverse, and has instability. Most of the generative models currently being studied generate data by using unsupervised learning to find the conditional probability p(x|z) given the latent variable z.

The generation model is a model that generates data based on a given probability distribution, and what is required to generate input data is to find a joint distribution as shown in Equation 1.

At this time, z is called a Latent Vector, and the goal of the generation model is to find out p(x), which is the probability distribution of the data. Therefore, p(x) is estimated by changing the parameter value and repeating the maximum likelihood estimation of the probability distribution.

Among the above-described generation models, a representative model is the Generative Adversarial Network ('GAN'), and GAN refers to the latent variable z as a characteristic of the input data. It will change. The aforementioned GAN is composed of two models serving as a generator and a discriminator. The discriminator learns to judge only real data as true, and the generator learns to generate fake data so that the discriminator cannot judge it as false. GAN's core idea is to reach the optimum point by simultaneously increasing the performance of both models through such hostile competition. However, the hostile generated neural network also has a problem that it is difficult to learn and is unstable, and mode collapsing is likely to occur in which the generated image is biased to a specific image that is difficult for the discriminator to distinguish. Various models are being proposed to solve these problems.

Meanwhile, semi-supervised learning ('SSL') is an intermediate learning method to compensate for the problems of supervised and unsupervised learning. As described above, semi-supervised learning constructs a model to learn by using both labeled and unlabeled data. In addition, there is a trend to more easily and quickly solve supervised learning problems such as classification and regression through this.

Accordingly, the present invention proposes a system and method capable of constructing a hostile generation model based on quasi-supervised learning with improved performance by combining hostile generation models.

Korean Registered Patent Publication No. 10-1925913 Korean Patent Application Publication No. 10-2019-0078710

An object of the present invention for solving the above-described problem is to provide a system and method capable of constructing a new type of hostile generation model with improved performance by combining hostile generation models based on quasi-supervised learning.

The hostile generation model building system according to the first aspect of the present invention for achieving the above-described technical problem comprises: a generator generating an image from an input random latent variable; A discriminator that receives the image generated by the creator as an input and determines whether it is real or fake; And a classifier for providing an output value by classifying using the result of the discriminator. And,

The discriminator is characterized in that it is configured by learning to optimize the discriminant's first loss function and the discriminator's second loss function for minimizing the Earth Mover's distance (EM distance) according to quasi-supervised learning, respectively, using learning data, and the The constructor is constructed by learning to optimize the third loss function of the generator of the basic hostile generation model and the fourth loss function of the generator for minimizing the EM distance, respectively.

In the hostile generation model building system according to the first feature described above, it is preferable that the learning data of the discriminator are labeled real data, unlabeled real data, and fake data generated by the creator.

In the hostile generation model construction system according to the first feature described above, the first loss function of the discriminator is a loss function based on labeled data according to supervised learning and a loss function due to unlabeled data according to unsupervised learning. It is preferably made up of the sum of.

In the hostile generation model construction system according to the first characteristic described above, the discriminator ?? It is preferable that the generator is composed of multiple layers of convolutional neural networks.

In the hostile generation model construction system according to the first feature described above, the classifier is preferably configured with a Softmax function, and is configured to classify and output the type of class and the presence or absence of a fake with respect to the result of the discriminator.

A method of constructing a hostile generation model based on quasi-supervised learning according to a second feature of the present invention is a generator that generates an image from a random latent variable and an image generated by the generator as inputs to determine whether it is real or fake. A method for constructing a hostile generation model having a discriminator to determine, comprising: (a) obtaining a first loss function for the discriminator according to quasi-supervised learning, and learning the first loss function to be optimized; (b) obtaining a second loss function for minimizing the Earth Mover's distance (EM distance) with respect to the discriminator, and learning the second loss function to be optimized; (c) obtaining a third loss function according to an existing hostile generation model with respect to the generator, and learning the third loss function to be optimized; (d) obtaining a fourth loss function for minimizing the EM distance with respect to the generator, and learning the fourth loss function to be optimized; (e) classifying the result of the discriminator using a classifier and providing an output value; It is equipped with.

In the method of constructing a hostile generation model according to the second feature described above, in steps (a) and (b), the discriminator inputs labeled real data, unlabeled real data, and fake data generated by the creator, It is desirable for the discriminator to learn using the input data.

In the method of constructing a hostile generation model according to the second feature described above, in step (a), the first loss function is preferably a sum of a loss function due to labeled data and a loss function due to unlabeled data. .

In the method for constructing a hostile generation model according to the second feature described above, it is preferable that the discriminator and the generator are composed of a plurality of layers of convolutional neural networks.

In the method for constructing a hostile generation model according to the second characteristic described above, in step (e), the classifier for classifying the result of the discriminator is configured with a softmax function, and is configured to classify and output the type of class and the presence or absence of fakes. desirable.

In the method for constructing a hostile generation model according to the second feature described above, the learning of steps (a) and (b) for the discriminator is performed a plurality of times, and steps (c) and (d) for the generator It is preferable that the process of learning is repeated once.

The system and method for constructing a hostile generation model with improved performance according to the present invention having the above-described configuration combines SGAN and WGAN to provide a newly constructed hostile generation model.

Since the system according to the present invention is based on a semi-supervised learning method, the learning speed is increased compared to the unsupervised learning method because it helps the discriminator to learn using the labeled real data.

In addition, the system according to the present invention can solve the Mode Collapsing problem according to the SGAN by applying the convolutional neural network structure of the Deep Convolution GAN to the SGAN.

In addition, the system according to the present invention applies a method of minimizing the EM distance according to the WGAN, so that even if the size of the data increases, the stability of learning can be improved.

1 is a schematic diagram showing a basic hostile generation model (GAN).
2 is a schematic diagram showing the structure of the generator of the Deep Convolution GAN.
3 is a schematic diagram showing the structure of a Semi-Supervised GAN (SGAN).
4 shows an algorithm for the learning process of Semi-Supervised GAN (SGAN).
5 is a schematic diagram showing the structure of a new type of GAN model based on semi-supervised learning according to a preferred embodiment of the present invention.
6 is a flowchart sequentially showing a method of constructing a hostile generation model based on quasi-supervised learning according to a preferred embodiment of the present invention.
7 is a diagram illustrating an algorithm for constructing a GAN combination model in a method of constructing a hostile generation model according to a preferred embodiment of the present invention.

The method and system for constructing a new form of hostility generation model according to the present invention improves learning speed and solves the problem of mode collapsing by optimizing the loss function of SGAN and the loss function of WGAN, respectively, and also improves the stability of learning. It is characterized by one.

Hereinafter, a new GAN model system based on semi-supervised learning according to a preferred embodiment of the present invention and a method of building the model will be described in detail with reference to the accompanying drawings.

1 is a schematic diagram showing a basic hostile generation model (GAN). Referring to FIG. 1, as Vanila GAN (VGAN), which is a basic hostile generation model, the basic hostile generation neural network is composed of two models serving as a generator ('G') and a discriminator ('D'). do. The constructor creates an image from a random latent variable z, whose purpose is to classify the output passing through the discriminator as a real image. On the other hand, the discriminator plays a role of discriminating the image created by the creator as an input, and its purpose is to accurately distinguish whether the input image is real or fake. Equation 2 is the objective function of GAN.

는

에 속하는 x라는 데이터들의 기대값이 된다. D(x)는 x에 대한 판별기이며, G(z)는 잠재변수 z에 대한 생성기이다. Here, V(G,D) is a value function that solves minmax for variables D and G, and E is the expected value.

Is

It becomes the expected value of the data x belonging to. D(x) is the discriminator for x, and G(z) is the generator for the latent variable z.

의 확률분포이고, x는 그 중 샘플링 데이터이다. 판별자 D는 출력이 실제 데이터가 들어오면 1에 가깝게 확률을 추정하고, 생성자 G가 만들어 낸 가짜 데이터가 들어오면 0에 가깝게 한다. 따라서, 로그를 사용했기 때문에 실제 데이터라면 최댓값인 0에 가까운 값이 나오고, 가짜 데이터라면 무한대로 발산하기 때문에 V(D,G)를 최대화하는 방향으로 학습하게 된다. 생성자 G 부분인 오른쪽 부분이 V(D,G)를 최소화하는 관점에서 보면,

는 보통 정규분포로 사용하는 임의의 노이즈 분포이고 z는 노이즈 분포에서 샘플링한 것이다. 이 입력을 생성자 G에 넣어 만든 데이터를 판별자 D가 진짜로 판별하면

이기 때문에,

는 1이며, D가 가짜로 판별하면

는 0 이기 때문에

는 0에 가까운 최댓값이 나오게 될 것이다. 따라서, G는 V(D,G)에 있어서 G는 이를 최소화하는 방향으로 D는 최대화하는 방향으로 간다. Looking at the objective function of GAN, from the point of view that D maximizes V(D,G)

Is the probability distribution of, and x is the sampled data among them. Discriminator D estimates the probability that the output is close to 1 when the actual data comes in, and closes to 0 when the fake data generated by the generator G comes in. Therefore, because the logarithm is used, a value close to 0, which is the maximum value for real data, is obtained, and if it is fake data, it is emitted to infinity, so the learning is performed in the direction of maximizing V(D,G). From the perspective of minimizing V(D,G), the right part, which is the constructor G part,

Is an arbitrary noise distribution normally used as a normal distribution, and z is a sampled from the noise distribution. If the discriminator D really determines the data created by putting this input into the constructor G

Because

Is 1, and if D is determined to be fake,

Because is 0

Will result in a maximum value close to zero. Therefore, G goes in the direction of minimizing it and D in the direction of maximizing it in V(D,G).

On the other hand, DCGAN (Deep Convolution GAN) is a development of the above-described VGAN structure, and applies a convolutional neural network to VGAN, and is the basis of the GAN model. DCGAN applied the following five methods to VGAN: (1) the maximum pooling layer was removed and the size of the feature map was adjusted using convolution, (2) batch normalization was applied, and (3) a fully connected hidden layer was used. Removed, (4) tanh is used as the output activation function of the constructor, relu is used for the rest of the layers, and (5) leakyrelu is used as the activation function of the discriminator.

2 is a schematic diagram showing the structure of the generator of the Deep Convolution GAN. Referring to FIG. 2, DCGAN constructs a DCGAN generator model through the above-described method, and as a result, it is possible to generate a much sharper image compared to VGAN, and a considerable number of mode collapsing problems occurring in the learning process are solved. .

The GAN model construction method according to the present invention creates a new model by combining a semi-supervised GAN (SGAN) using a DCGAN discriminator and generator structure and a loss function of Wassertein (WGAN), It is characterized in that the discriminator is learned using semi-supervised learning that uses both the data and the unlabeled fake data to be generated by the generator. The hostile generation model construction system according to the present invention can be implemented with a high-performance computer, etc., in which training data is stored in a database in advance, and the hostile generation model construction method is a storage device in the form of a program for building a hostile generation model. And the like, and executed and implemented by a computer or server of the hostile generation model building system.

Semi-Supervised GAN (SGAN) is an advanced GAN model based on the structure of DCGAN. Unlike the VGAN discriminator simply discriminating whether the image created by the creator is real or fake, the SGAN allows the discriminator to classify the class of input data and distinguish whether it is fake or not. Since the VGAN discriminator distinguishes only the real and the fake, it is output using the sigmoid function, but the SGAN must simultaneously distinguish the type of class and whether it is fake, so the softmax function is used to output it. In addition, the final purpose of VGAN is to obtain a constructor that generates a sophisticated fake image that the discriminator cannot distinguish, so the focus of learning is on the constructor, but SGAN is to learn a discriminator that can simultaneously distinguish the type of class and whether it is fake. It's the ultimate goal.

3 is a schematic diagram showing the structure of a Semi-Supervised GAN (SGAN). 3, Semi-Supervised GAN (SGAN) basically learns a discriminator through semi-supervised learning, and based on the latent variable z, the generator creates a fake image and the discriminator is labeled with this fake image. You will learn using all the real data. That is, the SGAN discriminator effectively learns the discriminator through semi-supervised learning using both labeled and unlabeled data, and through this, the creator can also learn through competition. The generator and discriminator basically use the structure of DCGAN.

In SGAN, since the discriminator must distinguish n classes and at the same time distinguish the image created by the creator as a fake, a total of n+1 classes must be classified. Assuming that x is fake data, the probability of each class of x can be expressed as in Equation 3.

Here, l is an abbreviation of logit, and since the classifier handles a vector of k+1 dimensions, the elements contained in the vector are l1, l2,… , it is defined as ln+1.

If x is real labeled data, it must be classified as one of the classes, so the probability of x can be expressed as shown in Equation 4.

Therefore, the loss function of the discriminator for each case is the same as in Equations 5 and 6.

Therefore, the final loss function of the discriminator D of the SGAN can be obtained by adding the aforementioned two loss functions as shown in Equation 7.

Meanwhile, the loss function of the generator G of the SGAN uses the same loss function as in the VGAN as shown in Equation 8.

By setting the loss function as above and learning the discriminator through semi-supervised learning, it is possible to obtain good results while learning the discriminator faster than the existing GAN.

4 shows an algorithm for the learning process of Semi-Supervised GAN (SGAN).

On the other hand, Wasserstein GAN (WGAN) is a model that seeks stability of learning by redefining the loss function in VGAN, and is characterized by using the Wassertein distance when setting the loss function. In general, VGAN uses JS divergence (Jensen-Shannon Divergence), whereas Wasserstein GAN (WGAN) uses Wassertein distance, which is also called Earth Mover's distance (EM distance). This is because EM distance interprets the concept of distance as the minimum cost to make a pile of some probability distribution shape have a different probability distribution shape. At this time, the cost is quantified by multiplying the amount of soil and the distance traveled. The reason why WGAN does not use the JS distance and uses the EM distance is that the sequence of probability distributions that do not converge at the JS distance sometimes converge at the EM distance.

5 is a schematic diagram showing the structure of a new type of GAN model based on semi-supervised learning according to a preferred embodiment of the present invention. 5, a new form of hostile generation model system according to the present invention is a combination of SGAN and WGAN models, and the generator and discriminator basically use the structure of DCGAN, and the convolutional neural network of layers 5 and 4, respectively. It consists of. The GAN model 10 according to the present invention includes a generator 20 and a discriminator 30, the generator and discriminator are composed of a convolutional neural network, and the discriminator further includes a classifier, so that the result value of the discriminator is the type of class. And the presence or absence of a fake is determined and output.

Hereinafter, a method of constructing a GAN model based on semi-supervised learning according to a preferred embodiment of the present invention will be described in detail with reference to FIG. 6. 6 is a flowchart sequentially showing a method of constructing a hostile generation model based on quasi-supervised learning according to a preferred embodiment of the present invention.

)의 값을 조정하면서 일부분의 데이터만을 레이블되고 나머지 데이터는 레이블없이 사용되도록 입력데이터가 구성된다. 따라서, 본 발명에 따른 결합 모델의 판별자로 입력되는 입력데이터는 레이블된 진짜 데이터, 레이블되지 않은 진짜 데이터 및 생성자가 생성한 가짜 데이터이다. In order to apply the semi-supervised learning method, the discriminator of the combined model according to the present invention is a label parameter (

While adjusting the value of ), the input data is configured so that only part of the data is labeled and the rest of the data is used without a label. Accordingly, the input data input to the discriminator of the combined model according to the present invention are labeled real data, unlabeled real data, and fake data generated by the creator.

When semi-supervised learning is performed, the learning speed is improved compared to the unsupervised learning method because it helps the discriminator to learn by using the labeled real data. However, at the beginning of learning, since the discriminator has a lower ability to distinguish the input image, the slope value has a large value and the learning proceeds well.However, if the learning speed is relatively faster than the generator and the classification performance increases, the Mode Collapsing problem occurs. I can. In this case, in order to further increase the stability of learning, the combined model according to the present invention is characterized by adding a method of minimizing the EM distance of the WGAN.

)는 수학식 9와 같은 레이블되지 않은 데이터를 이용한 비지도 학습에 따른 손실 함수와 수학식 10과 같은 레이블된 데이터를 이용한 지도 학습에 따른 손실함수의 합으로 이루어진다. 한편, WGAN의 판별자의 손실 함수(

)는 수학식 12과 같이 EM 거리를 최소화시키는 함수로 나타낼 수 있다. 따라서, 본 발명에 따른 GAN 결합 모델에 있어서, 판별자는 SGNA에 따른 손실함수(

)와 WGAN에 따른 EM 거리를 최소화하는 손실 함수(

)를 각각 최소화시킨다. The loss function of the GAN coupling model according to the present invention optimizes two types of loss functions through different optimization methods. First, the loss function of the discriminant according to semi-supervised learning (

) Is a sum of a loss function according to unsupervised learning using unlabeled data such as Equation 9 and a loss function according to supervised learning using labeled data such as Equation 10. Meanwhile, the loss function of the discriminator of WGAN (

) Can be expressed as a function that minimizes the EM distance as shown in Equation 12. Therefore, in the GAN coupling model according to the present invention, the discriminator is the loss function according to SGNA (

) And the loss function (

) Are minimized respectively.

)와 WGAN에 의한 EM 거리를 최소화하는 손실함수(

)를 각각 최소화시키게 된다. 수학식 13은 생성자는 SGAN에 의한 손실 함수(

)이며, 수학식 14는 WGAN에 의한 EM 거리를 최소화하는 손실함수(

)이다. In the GAN combination model according to the present invention, the generator is a loss function by SGAN (

) And the loss function (

) Are minimized respectively. Equation 13 shows that the generator is a loss function by SGAN (

), and Equation 14 is a loss function that minimizes the EM distance by WGAN (

)to be.

In summary, the GAN model construction method according to the present invention sets the loss function when there is a label and the loss function when there is no label for semi-supervised learning for the discriminator, and adds the values of the two loss functions. By giving, the loss function for quasi-supervised learning is obtained, and an optimization algorithm is applied to it. At this time, Adam Optimizer can be used as an optimization algorithm. In addition, a loss function for minimizing the EM distance according to the WGAN is set for the discriminator, and an optimization algorithm is applied thereto. In this case, the RMS Optimizer can be used as the optimization algorithm.

On the other hand, in the case of the generator, a loss function according to VGAN, which is a basic hostile generation model, is set and an optimization algorithm is applied. In this case, Adam Optimizer can be used as the optimization algorithm. In addition, the loss function according to the WGAN is set for the generator and an optimization algorithm is applied to it. In this case, the RMS Optimizer can be used as the optimization algorithm.

In the method of constructing a GAN model according to the present invention, the generator is trained once when the discriminator is trained multiple times, so that the discriminator is further for stability of learning in the process of clipping the weight to c to satisfy the condition of the WGAN loss function. You will be able to learn.

의 입력을 받아 동작한다. 먼저, 판별자 D를 5회 학습시킨 뒤 생성자 G를 1회 학습시켜야 하므로, t=0,…,5만큼 판별자 D를 우선적으로 학습시키게 된다. 각 수식들의 손실 함수는 전술한 바와 동일하다. 7 is a diagram illustrating an algorithm for constructing a GAN model according to the present invention in a method of constructing a hostile generation model according to a preferred embodiment of the present invention. Referring to Figure 7, the GAN model described above, first for each parameter

It receives the input of and operates. First, since the discriminator D needs to be trained 5 times and then the generator G must be trained once, t=0,... The discriminator D is learned preferentially by ,5. The loss function of each equation is the same as described above.

는 SGAN의 손실함수를 의미하며 앞의 두분은 unsupervised loss 이며, 뒤의 두 부분은 supervised loss로서, 이들의 합으로 이루어진다. 한편, 알고리즘의

는 WGAN의 손실함수를 의미한다. Algorithmic

Means the loss function of SGAN, the first two are unsupervised losses, and the latter two are supervised losses, which are summed up. Meanwhile, of the algorithm

Means the loss function of WGAN.

를 업데이트하게 된다. In the discriminator's learning process, the loss function of SGAN and the loss function of WGAN are variable using Adam optimizer and RMS optimizer, respectively.

Will be updated.

Next, the generator G is learned once, and proceeds in the same way as the discriminator learning method.

In the method for constructing a GAM coupling model according to the present invention, the discriminator uses a classifier last, and it is preferable to use a softmax function as the classifier. In this way, by using the softmax function as the classifier, the discriminator can determine and output the classification of the class and whether it is fake. Therefore, the discriminator of the GAN combination model according to the present invention learns through quasi-supervised learning, thereby distinguishing whether the data generated by the generator is real or fake, and classifying the data belonging to a label in the case of labeled real data. There will be.

On the other hand, the generator also uses a classifier at the end, and sigmoid is preferred as the classifier.

In the above, the present invention has been described with reference to its preferred embodiments, but these are only examples and do not limit the present invention, and those of ordinary skill in the field to which the present invention pertains will not depart from the essential characteristics of the present invention. It will be appreciated that various modifications and applications not exemplified above are possible in the range. And, differences related to these modifications and applications should be construed as being included in the scope of the present invention defined in the appended claims.

10: GAN model
20: constructor
30: discriminator
40: classifier

Claims (11)

A generator that generates an image from the input random latent variables;
A discriminator that receives the image generated by the creator as an input and determines whether it is real or fake; And
A classifier that classifies using the result of the discriminator and provides an output value;
And the discriminator is configured by learning to optimize the discriminator's first loss function and the second loss function of the discriminator for minimizing the Earth Mover's distance (EM distance) according to quasi-supervised learning, respectively, using learning data. And
The generator is a hostile generation model construction system based on quasi-supervised learning, characterized in that the generator is configured by learning to optimize the third loss function of the generator of the basic hostile generation model and the fourth loss function of the generator to minimize the EM distance. . The system of claim 1, wherein the learning data of the discriminator are labeled real data, unlabeled real data, and fake data generated by a generator. The method of claim 1, wherein the first loss function of the discriminator is
A system for constructing a hostile generation model based on quasi-supervised learning, characterized by consisting of a sum of a loss function based on labeled data according to supervised learning and a loss function due to unlabeled data according to unsupervised learning. The method of claim 1, wherein the discriminator is composed of a convolutional neural network of multiple layers,
The generator is a hostile generation model construction system based on quasi-supervised learning, characterized in that the generator is composed of a convolutional neural network of multiple layers. The hostile generation model based on quasi-supervised learning according to claim 1, wherein the classifier is configured with a Softmax function, and is configured to classify and output the type of class and the presence or absence of a fake with respect to the discriminator result. Build system. In a method for constructing a hostile generation model, comprising: a generator generating an image from a random latent variable and a discriminator determining whether it is real or fake by receiving the image generated by the generator as an input,
(a) obtaining a first loss function for the discriminator according to quasi-supervised learning, and learning the first loss function to be optimized;
(b) obtaining a second loss function for minimizing the Earth Mover's distance (EM distance) with respect to the discriminator, and learning the second loss function to be optimized;
(c) obtaining a third loss function according to an existing hostile generation model with respect to the generator, and learning the third loss function to be optimized;
(d) obtaining a fourth loss function for minimizing the EM distance with respect to the generator, and learning the fourth loss function to be optimized;
(e) classifying the result of the discriminator using a classifier and providing an output value;
A method of constructing a hostile generation model based on quasi-supervised learning, characterized in that it comprises a. The method of claim 6, wherein steps (a) and (b) are
The discriminator is a method of constructing a hostile generation model based on quasi-supervised learning, characterized in that the labeled real data, the unlabeled real data, and the fake data generated by the creator are input, and the discriminator learns using the input data. . The method of claim 6, wherein step (a)
The first loss function is a method of constructing a hostile generation model based on quasi-supervised learning, characterized in that the first loss function is composed of a sum of a loss function due to labeled data and a loss function due to unlabeled data. The method of claim 6,
The method of constructing a hostile generation model based on quasi-supervised learning, wherein the discriminator and the generator are composed of a plurality of layers of convolutional neural networks. The method of claim 6, wherein the step (e),
The classifier for classifying the result of the discriminator is composed of a softmax function, and is configured to classify and output the type of class and the presence or absence of fakes. The method of claim 6, wherein the learning of steps (a) and (b) for the discriminator is performed multiple times, and the learning of steps (c) and (d) for the generator is performed once. A method of constructing a hostile generation model based on quasi-supervised learning, characterized in that it is repeated.

Images