What are the advantages of generating a confrontation network (GAN) compared to traditional training methods? (i)

Author: Yuan Feng
Link: https://www.zhihu.com/question/56171002/answer/148593584
Source: Know
Copyright belongs to the author. Commercial reprint please contact the author to obtain authorization, non-commercial reprint please indicate the source.

Since 2014, when Ian Goodfellow put forward the concept of creating a confrontation network (GAN), the generation of confrontation networks has become a hot research hotspot for academia, and Yann LeCun is called "the most exciting idea in the field of machine learning in the last decade". A brief introduction to generating a confrontation network is as follows , training a generator (generator, or g) to generate realistic samples from random noises or potential variables (latent Variable) while training a discriminator (discriminator, abbreviation D) to identify real data and generate data, both at the same time, Until a Nash equilibrium is reached, the data generated by the generator is no different from the real sample, and the discriminator does not correctly distinguish between generating data and real data. The structure of Gan is shown in Fig. 1. <img src="https://pic1.zhimg.com/v2-61e1b1a6d1feb23a7d3c52966d11be08_b.png" data-rawwidth="759" data-rawheight="420" Class="origin_image zh-lightbox-thumb" width="759" Data-original="https://pic1.zhimg.com/v2-61e1b1a6d1feb23a7d3c52966d11be08_r.png">

Figure 1. Building a basic architecture against the network

In the past two years, the academic circles have put forward a variety of Gan varieties, such as conditional generation countermeasure network (Cgan), Information generation countermeasure Network (Infogan) and deep convolution generation confrontation Network (Dcgan), and the following figure 2 comes from last year's image-to-image translation With Conditional adversarial Nets paper, we can see that Gan has been introduced into a variety of previous deep neural network tasks, such as restoring the original image from the segmented image (the first pair in the upper left corner), coloring the Black-and-white picture (the first pair in the upper-right corner), coloring the texture ( In the lower right corner, in addition, Gan can also do image super resolution, dynamic scene generation, and so on more applications of Gan see another blog in-depth learning in the field of computer vision in the forefront of progress.

<img src="https://pic4.zhimg.com/v2-94c48268909fce330c553eb2f5e214c3_b.png" data-rawwidth="1396" data-rawheight="524" Class="origin_image zh-lightbox-thumb" width="1396" Data-original="https://pic4.zhimg.com/v2-94c48268909fce330c553eb2f5e214c3_r.png">

Figure 2. Images to image translation

Think carefully, these tasks, in fact, are traditional deep neural network can do, such as from the Encoder (Autoencodor) and convolution deconvolution structure can be done, we can't help but think, gan compared with the traditional depth of neural network, its advantages in where? Some time ago, I have been more puzzled, The information that can be found in Chinese is the final summary of Ian Goodfellow's thesis on the generation of the Antagonism network (GAN), as follows: The Advantage model uses only the reverse propagation, but does not need to infer the implicit variable in the Markov chain training without the theory, As long as a differentiable function can be used to build D and G, because the parameter updates that can be combined with a deep neural network to make a depth-generated model g are not directly derived from the data sample, but rather from reverse propagation from D (which is also the biggest difference from the traditional method) disadvantage The explanatory difference is poor, the distribution of the generated model PG (g) Without explicit expression is difficult to train, D and G need good synchronization, such as D update K and G update once

It's just a simple explanation, and luckily, I found two similar questions on Quora, and what advantages does Gan have over other models, that only one person answered, and luckily, he was the inventor of Ian Goodfellow,gan, His signature on the Quora is "I invented generative adversarial networks". And the other question is, what are the pros and cons of Gans? The great God of conscience Goodfellow also did the answer! I translate his two answers as follows:

Split Line
————————————————————————————————— – original question 1:What is the advantage of generative adversarial networks compared with OT Her generative models?

What are the advantages of generating a network against the other generation models?
Ian Goodfellow replied:
—————————————————————————————————–
Compared to all other models, I think from the actual results, Gan seems to produce a better sample of the GAN framework can train any build network (theoretically, however, in practice, it is difficult to use enhanced learning to train the generator with discrete output), most other architectures require that the generator has some specific function forms, Just as the output layer must be Gaussian. In addition all other frameworks require the generator to be a total of 0 weights (put Non-zero mass everywhere), however, Gans can learn a model that produces sample points only near real data (the neural Network layer) (Gans can learn models That generate points is only on a thin manifold that goes near the data.) There is no need to follow any kind of factor decomposition to design the model, all generators and discriminator can work properly compared to PIXELRNN, Gan sample production time is shorter, Gans a sample, but Pixelrnns need one pixel pixel to produce samples; Compared to VAE, Gans has no variational lower bound, if the discriminator is well trained, the generator can perfectly learn the distribution of the training samples. In other words, the Gans is asymptotically consistent, but VAE is biased compared to the depth of the Boltzmann machine, the Gans does not have a variational lower bound, and there is no tricky division function, The sample is generated one at a time instead of the repeated application of the Markov chain to generate the sample generated by the Gans rather than the repeated application of the Markov chain Gsns. Compared to the nice and real nve,gans, there is no limit on the size of the potential variable (the input value of the generator);
To tell you the truth, I think the other methods are also great, they have a corresponding advantage compared to Gans.
————————————————————————————————— – original question 2: What are the pros and cons of using generative adversarial networks (a Ty PE of neural network)?

What are the pros and cons of generating a confrontation network (a neural network)?
It is known that Facebook has developed a means of generating realistic-looking via a images neural. They used "GAN" aka "Generative Adversarial Networks". Could this is applied generation of other things, such as audio waveform via RNN? Why or Why not?
Facebook has developed a way to generate realistic images based on neural networks, using Gan, also known as a generation-versus-network, that can be used to generate other things, such as generating audio waveforms through RNN, right?
Ian Goodfellow replied:
—————————————————————————————————–
Advantages Gans is a method of training classifiers in a semi supervised manner, which can refer to our Nips paper and corresponding code. When you don't have a lot of tagged training sets, you can just use our code without making any changes, Usually this is because you don't have too many tag samples. I have also recently succeeded in using this code to write a paper with the Google Brain Department in deep learning privacy Gans can produce samples faster than a fully visible belief network (nade,pixelrnn,wavenet, etc.) Because it does not need to generate different data in the sample sequence. Gans does not need Monte Carlo estimates to train the web, and people often complain that Gans training is unstable and difficult to train, but they are much simpler than training Boltzmann machines that rely on Monte Carlo estimates and logarithmic distribution functions. Because the Monte Carlo method is not effective in high-dimensional space, The Boltzmann machine has never been extended to tasks like imgenet. Gans at least after training on imagenet, you can learn to draw some really decent dogs. Gans did not introduce any deterministic bias (deterministic bias) compared to the variational self encoder, The variational approach introduces a deterministic bias because they optimize the lower bound of the logarithm, rather than the likelihood itself, which seems to cause the vaes generated instances to be more obscure than the Gans. Compared to nonlinear ICA (Nice, real nve, etc.), Gans does not have any specific dimensions for the potential variable entered by the generator or requires that the generator be reversible. The process of generating an instance of the Boltzmann machine and Gsns,gans is only required to run the model once rather than iterating over many times in the form of a Markov chain.

Disadvantage training gan requires a Nash equilibrium, sometimes it can be done with a gradient descent method, and sometimes not. We have not found a good way to achieve Nash equilibrium, so it is unstable to train gan compared to vae or PIXELRNN, But I think in practice it is still much more stable than training the Boltzmann machine. It's hard to learn to generate discrete data, like text compared to the Boltzmann machine, Gans it's hard to guess another pixel value based on a pixel value, Gans is born to do one thing, that is to create all the pixels at once, you can use Bigan to correct this feature, It allows you to use the Gibbs sample to guess the missing value like a Boltzmann machine,
I talked about it 20 minutes before the Berkeley class. Course links, Tubing video, please bring a ladder ~
——————————————————————————————— --–
above is the translation of Ian Goodfellow's original answer, please correct me if there is something wrong with the translation. The
was recently seen in a paper in which Gan is a picture-generating document. The authors suggest that Gan can obtain better generalization results than normal convolution deconvolution. I don't know if I'm using discriminator D as a monitor, when D's precision is close to 50%, stop training to prevent the generator from fitting.
Gan There's something else I'd like to know about, and then I'll add it to the article. Thank you ~