[paper notes] - gan-2014-nips

subject ：Generative Adversarial Nets

Generating antagonism neural network GAN The first paper .

The author of the paper is “ The father of generative antagonism ”Ian Goodfellow And the Turing prize winner Youshua Bengio.

GAN In recent years, it has become a hot research field of artificial intelligence and deep learning .GAN Widely used in image generation 、 Style transfer 、AI art 、 Black and white old photo coloring repair . You can use GAN Turn photos into oil paintings 、 Wild horses turn into zebras 、 Night turns into day , The cat with simple strokes turns into a real cat , Cool and fun applications such as turning blurred images into high-definition images .

EID：arXiv:1406.2661
DOI：10.5555/2969033.2969125

Time ：2014-06-10 Upload on arxiv
Periodical ：2014 27th International Conference on Neural Information Processing Systems（NIPS）
Institutions ： Canada - University of Montreal

Thesis link ：https://arxiv.org/abs/1406.2661
Code link ：http://www.github.com/goodfeli/adversarial
Author URI ：Ian Goodfellow --- Research Page

GAN Lab Webpage ：GAN Lab: Play with Generative Adversarial Networks in Your Browser!
OpenMMLab Image generation open source algorithm library MMGeneration：https://github.com/open-mmlab/mmgeneration

The basic principle ： The game between generator network and discriminator network .

GAN It is mainly composed of two networks , Generator network (Generator) And discriminator networks (Discriminator), Through the mutual game between the two networks , So that the generator network can finally learn the distribution of input data , This is the same. GAN What you want to achieve -- Learn the distribution of input data .

D It's a discriminator , Responsible for the input of real data and by G Judge the generated false data , Its output is 0 and 1, That is, it is essentially a binary classifier , The goal is that the input is real data, and the output is 1, Input of false data , The output is 0;

G It's a generator. , It receives a random noise , And generate images .

During the training ,G The goal is to generate as real data as possible to confuse D, and D It's about putting G The generated pictures can be identified , In this way, the two are playing games with each other , The ultimate goal is to achieve a balance , That's Nash equilibrium .

GAN A network model ：

Generative antagonistic network (GAN) Think of the training process as a game between two independent networks ： The discriminator attempts to classify the samples as samples from the real distribution p(x) Or model distribution P(x), Every time the discriminator notices the difference between the two distributions , The generator will slightly adjust its parameters to make it disappear , Until the last （ Theoretically ） The generator accurately reproduces the real data distribution , So that the discriminator cannot find a difference .

Iterative gradient descent modification of the input image , Generate some noise and add it to the original image , Make the output result of neural network refer to deer as horse .

1. GAN Model design is a differentiable function .
2. The generation process is estimated through a confrontation process .
3. Training at the same time 2 A model ： Generate a model G Used to capture data distribution , A discriminant model D Used to estimate the probability of training data .

Generate models ：

 

Countermeasure network framework ：

G Try to generate D Samples that are difficult to distinguish from data “ cheating ”D.

1. Training K Secondary discriminator , Train once .
2. The discriminator guides the generator to generate a more realistic image .
3. The generator attempts to maximize the probability that the discriminator mistakenly believes its input to be true .
4. The generator does not fit directly with the original data , Instead, use the discriminator to train indirectly （ Avoid overfitting ）.
5. The generator is trained with the discriminator , Otherwise “the Helvetica scenario” The phenomenon of mode collapse （G Taste the sweetness , No longer evolve , The generated images are the same ）.
6. The generator does not really generate false data, but deceives the discriminator D（ give an example ： Kallima inachus ）

Max min function ：

The derivative is back propagated through the generation process , The generation process refers to the following observation functions : 

On the left ：maxD  Given G Find the V To maximize the D,maxG  Given D Find the V Minimized G;
the aforesaid ： When the discriminator inputs real data , The larger the output, the better ;
consequent ： When the discriminator inputs false data , The smaller the output, the better .

Optimal discriminator ：

The learning process ：

pD(data)： The probability that the discriminator discriminates the data as real data ;
Data distribution： The distribution of real data ;
Model distribution： The generator generates the distribution of data ;

1. Noise space Z（ Uniformly distributed random number ） The distribution probability of the generated false image and the true image under the discriminator pD(data).
2. In the inner loop of the algorithm ,D It is used for training to distinguish samples and data , Converging on the optimal discriminator D(x).
3. In the face of G After updating ,D Gradient guidance G(z) Flow to areas that are more likely to be classified as data .
4. After several steps of training , If G and D Have enough ability , At this point they will reach a point , At this point, neither can be improved , because p g=p data. At this time, the discriminator cannot distinguish the two distributions , because D(x)=1/2.

experimental result ：

chart 2： Visualization of model samples .

The fifth column ：GAN Generated data （ The generated data is very realistic ）

The sixth column ： and GAN Generate the closest training data , To prove that the model has not yet remembered the training set .

Input noise gradient , The generated image also follows the gradient .

A one-dimensional ： chart 3- Through the z The number obtained by linear interpolation between coordinates in space .

A two-dimensional ：Learned 2-D manifold of MNIST Visual track of

Quantitative likelihood results：

The fidelity of the data generated on two different data sets （ The larger the average, the better ）, mean value + variance , this paper Adversarial nets

GAN And the difference between confrontation samples ：

GAN： Iteratively modify the input image , Misleading classification neural network “ deliberately misrepresent ”;

Counter samples ： Only modify the input image , Not used for training network .

GAN and VAE（ Variational automatic encoder ） The difference between ：

GAN advantage ：

1. According to the actual results , They seem to produce better samples than other models （ The image is sharper 、 Clear ）.
2. GAN The model only uses back propagation , Without repeated sampling by Markov chain , Avoid the difficult problem of approximate calculation of probability .
3. There is no need to infer hidden variables during training
4. Theoretically , Any differentiable function can be used to construct D and G , Because it can be combined with depth neural network to make depth generative model
5. G The parameter update of is not directly from the data sample , Instead, it uses information from D Back propagation of
6. Compared with other generation models （VAE、 Boltzmann machine ）, Better samples can be generated
7. GAN It is a semi supervised learning model , You don't need too much labeled data for the training set ;
8. There is no need to follow any kind of factorization to design the model , All generators and discriminators work properly

GAN Existing problems ：

1. Solve non convergence （non-convergence） The problem of
2. It's hard to train ： Breakdown problem （collapse problem）： The generator starts to degrade
3. No need to model in advance , The model is too free and uncontrollable

Reference material ：

Back propagation ：Neural networks and deep learning

Generate models ：Generative Models

Reference blog ：A Short Introduction to Generative Adversarial Networks - Thalles' blog

The author's speech ：https://www.iangoodfellow.com/slides/2015_icml_gans_wkshp_invited.pdf

Reference blog ：https://blog.csdn.net/solomon1558/article/details/52549409

Intensive reading ： Generative antagonistic network GAN Intensive reading of the thesis of the first mountain _ Bili, Bili _bilibili

Zhihu explains ：[GAN Learning Series ] First time to know GAN - You know