深度学习理论与应用：生成对抗网络解析

需积分: 5 116 浏览量更新于2024-06-18 收藏 20.66MB PDF 举报

"《深度学习理论与应用》深入探讨了生成对抗网络在各种应用中的使用，强调了它们相比传统生成模型的显著进步。本书的主要目标是聚焦深度学习和生成对抗网络的前沿研究，包括为处理文本、图像和音频创建新的工具和方法。书中详细介绍了如何利用生成对抗网络（GAN）来生成符合应用需求的数据样本，涵盖了GAN在科学和技术领域的广泛应用，如计算机视觉、安全、多媒体和广告、图像生成、图像转换、文本到图像合成、视频生成、生成高分辨率图像以及药物发现等。" 深度学习是现代人工智能领域的一个关键分支，它主要通过模拟人脑神经网络的工作方式来学习数据的复杂表示。生成对抗网络（GANs）是深度学习中的一个重要概念，由两个神经网络构成：生成器（Generator）和判别器（Discriminator）。生成器的任务是根据随机噪声生成与训练数据相似的新样本，而判别器则试图区分生成器产生的假样本和真实数据。通过这两个网络之间的对抗性训练，GANs能够逐步提升生成样本的质量，达到以假乱真的效果。本书特色在于提供了一个全面的指南，详细介绍如何将GAN应用于图像和视频处理。例如，在图像生成方面，GAN可以用来创作艺术作品，模仿特定艺术家的风格，或者进行图像修复和增强。在图像翻译中，GAN可以实现不同类型的图像之间的转换，如将黑白图像转换为彩色，或者将夏天的景色转换为冬天的景色。此外，文本到图像的合成功能使GAN能够在给定文本描述时生成相应的图像，这在插图制作或虚拟现实应用中具有巨大潜力。在视频生成领域，GAN可以用于创造逼真的动态场景，这对于电影制作、游戏开发和虚拟现实体验来说具有革命性的意义。同时，GAN在高分辨率图像生成方面的应用，可以改善医学影像分析、遥感图像处理等领域，提供更清晰的细节视图。在药物发现中，GAN可以帮助快速生成可能的分子结构，加速药物研发过程。书中还包含水下图像增强的案例研究，这是一个具体的应用示例，展示了GAN如何改进因水体散射和吸收导致的图像质量下降问题，提高水下环境的视觉感知和分析能力。《深度学习理论与应用》通过丰富的实例和详尽的理论解释，为读者提供了深入理解并应用生成对抗网络的宝贵资源，无论是在学术研究还是实际工程中，都能为深度学习从业者带来启示和帮助。

1DOI: 10.1201/9781003203964-1

Generative Adversarial Networks and Its Use Cases

Chaitrali Sorde, Anuja Jadhav, Swati Jaiswal, Hirkani Padwad, and Roshani Raut

CONTENTS

1.1 Introduction ............................................................................................................................1

1.2 Supervised Learning .............................................................................................................2

1.2.1 Unsupervised Learning ............................................................................................ 3

1.3 Background of GAN .............................................................................................................4

1.3.1 Image-to-Image Translation .....................................................................................5

1.4 Difference between Auto Encoders and Generative Adversarial Networks ................7

1.4.1 Auto Encoders ............................................................................................................7

1.4.2 Generative Adversarial Networks .......................................................................... 7

1.5 Difference between VAN and Generative Adversarial Networks .................................7

1.6 Application of GANs ............................................................................................................8

1.6.1 Application of GANs in Healthcare .......................................................................8

1.6.2 Applications of Generative Models ........................................................................9

1.6.2.1 Generate Examples for Image Datasets ...................................................9

1.6.2.2 Generate Realistic Photographs ................................................................ 9

1.6.2.3 Generate Cartoon Characters ....................................................................9

1.6.2.4 Image-to-Image Translation ......................................................................9

1.6.2.5 Text-to-Image Translation .........................................................................9

1.6.2.6 Semantic-Image-to-Photo Translation ...................................................10

1.6.2.7 Photos to Emojis ........................................................................................ 10

1.6.2.8 Photograph Editing...................................................................................10

1.6.2.9 Face Aging ..................................................................................................10

1.7 Conclusion ............................................................................................................................ 10

References ......................................................................................................................................10

1.1 Introduction

Generative adversarial networks (GANs) have recently been a popular study area. “GANs

are the most exciting notion in machine learning in the last 10 years,” claimed Yann

LeCun, a deep learning legend, in a Quora post. According to Google Scholar, there are a

vast number of papers relevant to GANs. In 2018, there are approximately 11,800 papers

Generative Adversarial Networks and Deep Learning2

connected to GANs as well as 32 papers per day and more than one paper every hour

linked to it. A generator and a discriminator are the two models that make up a GAN.

Predictability minimization [1,2] is the most relevant work. Emami et al. [3,4] discuss the

relationship between predictability minimization and GANs. GANs have been the sub-

ject of multiple previous reviews due to their popularity and relevance. The following

is a summary of the differences between this and earlier studies. GANs for specic

applications: There have been surveys of GANs being used for specic applications such

as picture synthesis and editing [5], audio improvement and synthesis, and so on.

A GAN comprises of two contending neural organizations: a generative organization

known as a generator and assigned G, and a discriminative organization known as a dis-

criminator and stamped D, the two of which are enlivened by the two-player minmax

game. The generator network attempts to deceive the discriminator by making sensible

examples, while the discriminator attempts to recognize genuine and counterfeit examples.

A strategy can be utilized as the generator and discriminator organizations, as long as the

generator can become familiar with the circulation of the preparation information and the

discriminator can extricate the component to order the result.

A convolutional neural network or a recurrent neural network can be used as the dis-

criminator network, while a de-convolutional neural network can be used as the generator

network. As a result, GANs can be used to build multidimensional data distributions like

pictures. GANs have shown potential in a variety of difcult generative tasks, including

text-to-photo translation, picture generation, image composition, and image-to-image

translation [3]. GANs are a powerful type of deep generative model; however, they have a

variety of training issues, such as mode collapse and training instability. There are different

types of learning approaches in machine learning such as supervised and unsupervised

learning.

1.2 Supervised Learning

The PC is educated as a visual cue in administered learning. It takes what it has gained

from past information and applies it to current information to estimate future occasions.

Both the info and designated yield information are helpful in foreseeing future occasions

in this situation.

The input data is labeled or tagged as the correct response for accurate predictions.

Figure 1.1 shows the supervised learning approach.

Regulated learning, similar to all articial intelligence (AI) calculations, depends on pre-

paring. In Figure 1.1. the framework is provided named informational indexes all through

its preparation stage, which helps know the yield related to every particular information

esteem. The prepared model is then given test information, which is marked information

with the names stowed away from the calculation. The testing information is utilized to

decide how well the calculation performs on unlabeled information.

By consistently estimating the model’s results and tweaking the framework to draw

nearer to its objective precision, the regulated educational experience is moved along. The

degree of exactness still up in the air by two factors: the marked information accessible and

the calculation used.

Information from preparing should be adjusted and scrubbed. Information researchers

should be cautious with the information the model is prepared on the grounds that trash or

Generative Adversarial Networks 3

copy information will slant the AI’s perception. The variety of the information inuences

how really the AI performs when confronted with new situations; assuming that the prep-

aration information assortment contains inadequate examples, the model will vacillate and

neglect to give dependable reactions. Order and relapse are the two fundamental results of

administered learning frameworks.

1.2.1 Unsupervised Learning

Unsupervised learning, on the other hand, is a technique for teaching computers to use

data that has not been classied or labeled. It means that no preparation information is

accessible and the machine is customized to learn all alone. With no earlier information on

the information, the machine should have the option to group it.

The objective is to open the machines to huge measures of different information and

allow them to gain from it to uncover already obscure bits of knowledge and reveal stowed

away examples. Accordingly, solo learning calculations don’t necessarily create unsur-

prising outcomes. Rather, it gures out what makes the given dataset novel or interesting.

It is important to program the machine to learn all alone. Both organized and unstructured

information should be perceived and examined by the PC. Solo learning calculations can

deal with more complicated handling undertakings than regulated learning frameworks.

Besides, one method of scrutinizing AI is to utilize solo learning. Here’s an accurate illus-

tration of unsupervised learning in Figure 1.2.

The objective of solo learning is for calculations to identify designs in preparing informa-

tional collections and order input things in view of those examples. To separate applicable

data or highlights, the calculations dissect the informational collections’ central construc-

tion. Accordingly, by searching for joins between each example and info thing, these

FIGURE 1.2

Unsupervised learning approach.

FIGURE 1.1

Supervised learning approach.

Generative Adversarial Networks and Deep Learning4

calculations ought to have the option to create characterized yields from unstructured

information sources.

Informational indexes including creature photographs, for instance, could be given to

calculations. The animals may then be isolated into bunches in light of their fur, scales, and

plumes. It might order the photos into perpetually nitty gritty subgroups as it gures out

how to perceive differentiations inside every class.

The calculations achieve this by nding and recognizing designs; however, design

acknowledgment happens in solo learning without the framework being taken care of

information that trains it to recognize well-evolved creatures, shes, and birds, or among

canines and felines in the warm-blooded animal classication, for instance. Grouping is a

well-known unaided learning subject [7].

1.3 Background of GAN

GAN is a phenomenal generative model. In any case, the rst GAN model has numerous

issues, like the poor diversity, vanishing gradient, and training difculty [8]. Numerous

endeavors have been committed to getting better GANs through various enhancement

strategies. Hence, beginning around 2014, speculations and articles connected with GAN

have arisen in an interminable stream, and numerous new GAN-based models have been

proposed to work on the security and nature of the produced outcomes.

A generative adversarial network (GAN) is a sort of AI proposed by Ian Goodfellow

and his associates in a course named “Generative Adversarial Nets” [6] in June 2014. In

a game, two brain networks go up against each other. This cycle decides how to produce

new data with comparable experiences as the planning set given an arrangement set.

For instance, a GAN prepared on pictures can create new photographs that are cursorily

real to human spectators and have an assortment of helpful properties. GANs, which

were rst introduced as a kind of generative model for independent learning, have like-

wise demonstrated helpful for semi-directed learning, support learning, and completely

administered learning.

The “backhanded” prepping through the discriminator, which is also being updated

gradually, is at the heart of a GAN’s central concept. This simply means that the generator

is attempting to deceive the discriminator rather than limiting the distance to a certain

image. This allows the model to learn without being supervised.

Setting two calculations in opposition to one another started with Arthur Samuel, a

noticeable scientist in the eld of software engineering who’s credited with promoted the

expression “machine learning.” While at IBM, he concocted a checkers game—the Samuel

Checkers-playing Program—that was among the rst to effectively self-learn, to some

degree by assessing the shot at each side’s triumph at a given position.

In 2014, Goodfellow and his associates provided the main working execution of a gen-

erative model dependent on adversarial networks.

Goodfellow has frequently expressed that he was roused by noise contrastive

assessment, a method of learning information dissemination by looking at it against a

noise distribution, that is, numerical capacity addressing distorted or corrupted informa-

tion. The noise contrastive assessment involves a similar loss function as GANs—at the

end of the day, a similar proportion of execution concerning a model’s capacity to expect

anticipated results.

Generative Adversarial Networks 5

Once more, GANs comprise of two sections: generators and discriminators. The gen-

erator model produces engineered models (e.g., images) from arbitrary noise sampled

utilizing a distribution model, which alongside genuine models from a preparation infor-

mational collection are taken care of to the discriminator, which endeavors to recognize the

two. Both the generator and discriminator work on in their particular capacities until the

discriminator can’t perceive the genuine models from the integrated models with better

than the half precision expected of possibility.

GANs are algorithmic plans that utilize two brain afliations, setting one as opposed

to the accompanying to make new, created cases of information that can be pass with

genuine information. They are utilized regularly in picture age, video age and voice age.

Generative adversarial networks embraces up a game-hypothetical strategy, rather than

standard brain organizations. The organizations advance out a viable method for making

from a preparation appropriation through a two-player game. The two enemies are gen-

erator and discriminator. These two foes are in reliable ght all through the arranging cor-

respondence. Since an ill-disposed learning technique is taken on, we genuinely need not

care about approximating obstinate thickness capacities.

One brain organization, called the generator, delivers new data events, while the other,

the discriminator, evaluates them for validness; for instance, the discriminator closes

whether each instance of data that it reviews has a spot with the certiable planning

dataset or not.

As you can recognize from their names, a generator is used to make genuine looking

pictures and the discriminator’s liability is to perceive which one is phony. The

components/enemies are in a consistent battle as one (generator) endeavors to deceive the

other (discriminator), while various takes the necessary steps not to be deceived. To make

the best pictures, you will require a marvelous generator and a discriminator. This is since,

in such a case that your generator isn’t sufcient, it will not at any point have the choice to

deceive the discriminator and the model will not at any point join together. Accepting the

discriminator is terrible, then, picture which has no perceptible pattern will similarly be

named authentic and in this manner your model never plans and consequently you never

convey the best outcome. The information and unpredictable upheaval can be a Gaussian

appointment and characteristics can be inspected from this course and dealt with into the

generator association and a picture is created. This made the picture differentiated and a

genuine picture by the discriminator and it endeavors to perceive if the given picture is

phony or certied.

1.3.1 Image-to-Image Translation

GAN-based methods have been extensively used in picture to-picture interpretation and

conveyed drawing in results. Inpix2pix, contingent GAN (cGAN), was used to get to know

a preparation from an info picture to a result picture; cGAN gains a prohibitive generative

model using matched pictures from source and target regions. CycleGAN was utilized

for picture to-picture interpretation endeavors in the nonparticipation of joined models. It

takes in a preparation from a source region A to a goal space B by introducing two cycle-

consistency disasters. Likewise, DiscoGAN and DualGAN use an independent learning

approach for picture-to-picture interpretation reliant upon unaided data, yet with different

misfortune capacities. Consonant GAN utilizes for solo picture to-picture understanding

presents spatial smoothing to maintain unsurprising mappings during translation.

Image-to-image interpretation is to take in a preparation between pictures from a source

region and pictures from an objective region and has various applications including picture

剩余222页未读，继续阅读

死磕代码程序媛

粉丝: 110
资源: 315

深度学习理论与应用：生成对抗网络解析

Deep Learning_ Fundamentals, Theory and Applications.pdf

Deep Learning.pdf

Applications for Deep Learning.pdf

Manifold learning theory and applications

Python Machine Learning. Machine Learning and Deep Learning with Py, scikit

源码Python Machine Learning. Machine Learning and Deep Learning with Py, scikit

Deep Belief Nets in C and CUDA C Volume 1.pdf

MIT.Press.Deep.Learning.2016

deep learning

Deep Learning

最新资源