Research progress of generative anti-network gan (i.)

"Preface"
This paper first introduces the generative model, and then focuses on the research and development of generating antagonistic network (generative adversarial networks) in the Generative model (generative Models). According to Gan main thesis, gan applied paper and gan related papers, the author sorted out 45 papers in recent two years, focused on combing the links and differences between the main papers, and revealing the research context of the generative antagonism network.
the papers involved are:
[1] Goodfellow Ian, Pouget-abadie J, Mirza M, et al. generative adversarial nets[c]//advances in neural information Proces Sing Systems. 2014:2672-2680.
[2] Mirza M, Osindero S. Conditional generative adversarial nets[j]. Computer Science, 2014:2672-2680.
[3] Denton E L, Chintala S, Fergus R. Deep generative Image Models using a Laplacian Pyramid of adversarial networks[c]//a Dvances in neural information processing systems. 2015:1486-1494.
[4] Radford A, Metz L, Chintala S. Unsupervised representation learning with deep convolutional generative adversarial net WORKS[J]. ARXIV preprint arxiv:1511.06434, 2015.
[5] Im D J, Kim C D, Jiang H, et al generating images with recurrent adversarial networks[j]. ArXiv preprint arxiv:1602.05110, 2016.
[6] Larsen a B L, Sønderby S K, Winther O. autoencoding beyond pixels using a learned similarity metric[j]. ARXIV preprint arxiv:1512.09300, 2015.
[7] Wang X, Gupta A. Generative Image Modeling using Style and Structure adversarial networks[j]. ArXiv preprint arxiv:1603.05631, 2016.
[8] Chen X, Duan Y, Houthooft R, et al infogan:interpretable representation learning by information maximizing Generativ E adversarial nets[j]. ArXiv preprint arxiv:1606.03657, 2016.
[9] Kurakin A, Goodfellow I, Bengio S. Adversarial examples in the physical world[j]. ArXiv preprint arxiv:1607.02533, 2016.
[Ten] Odena A. semi-supervised learning with generative adversarial networks[j]. ArXiv preprint arxiv:1606.01583, 2016.
[One] Springenberg J T. Unsupervised and semi-supervised learning with categorical generative adversarial Networks[J]. ARXIV preprint arxiv:1511.06390, 2015.1. Built-in model 1.1 Overview

Machine learning methods can be divided into generative methods (generative approach) and discriminant methods (discriminative approach), and the models learned are called generative models (generative Model) and discriminant (discriminative models) [1 Hangyuan Li]. The generation method learns the joint probability distribution of samples and labels by observing data p (X, Y), and the trained model can generate new data that conforms to the sample distribution, which can be used for supervised learning and unsupervised learning. In supervised learning tasks, conditional probability distributions p (x, y) are obtained from the combined probability distribution P (y|) according to the Bayesian formula. X) to get a predictive model, typical models are naive Bayes, mixed Gaussian model, and hidden Markov model. Unsupervised generation model describes the distribution characteristics of sample data by learning the essential characteristics of real data, and generates new data similar to training samples. The parameters of the generated model are much smaller than the amount of training data, so the model can discover and effectively visualize the nature of the data, thus generating the data. The generative model occupies the primary position in unsupervised deep learning, and can be used to capture high-order correlations of observed or visible data without the target class tag information. The depth generation model can efficiently generate samples by sampling from the network, such as restricted Boltzmann machines (Restricted Boltzmann machine, RBM), depth belief network (deep Belief network, DBN), depth Boltzmann machine (deep Boltzmann machine, DBM) and generalized de-noising self-encoder (generalized denoising autoencoders). The most popular generative models in recent two years are divided into three methods [OpenAI first study]:Generate a Confrontation network (gan:generative adversarial Networks)
The two-person zero-sum game in Gan-inspired game theory, pioneered by [Goodfellow et al, NIPS 2014], consists of a generation model (generative model G) and a discriminant model (discriminative model D). The model captures the distribution of sample data, and the discriminant model is a two classifier, which is the discriminant of whether the input is real data or a generated sample. The optimization process of this model is a "two-dollar Minimax game (minimax two-player game)" problem, the fixed side of training, updating the parameters of another model, alternating iterations, so that the other side of the error maximization, and finally, G can estimate the distribution of sample data.variational Self-encoder (vae:variational autoencoders)
The problem is formalized in the framework of the probabilistic graphical model (probabilistic graphical models)-in the probabilistic graph model, we maximize the minimum (lower bound) of the logarithmic likelihood of the data.Autoregressive Model (autoregressive models)
Pixelrnn Such an autoregressive model trains the network by modeling the conditional distribution of each individual pixel, given a previous pixel (left or upper). This is similar to inserting pixels of an image into Char-rnn, but the RNN runs simultaneously in the horizontal and vertical directions of the image, not just the 1D sequence of characters.1.2 Classification of generative models [heavyweight | Yoshua Bengio Deep Learning Summer Course]

full observation model (Fully observed Models)
The model directly observes the data without introducing any new non-observable local variables. This type of model can directly compile the relationship between the observer points. For the directional graph model, it is easy to expand into large models, and since the logarithm probability can be computed directly (no approximation is required), parameter learning is also easy. For non-directional models, parameter learning is difficult because we need to calculate normalization constants. The generation in the full-watch model can be slow. The following figure shows the different full-watch generation models [images from Shakir Mohamed's presentation]:

Transform Model (transformation Models)
The model uses a parameterized function to transform a non-observed noise source. It's easy to do (1): sample from these models (2): Only the expected value is calculated without knowing the final distribution. They can be used in large classifiers and convolutional neuron networks. However, it is difficult to maintain reversibility with these models and extend them to general data types. The following figure shows the different transformation generation models [images from Shakir Mohamed's presentation]:

implicit variable model (latent Variable Models)
These models introduce a non-observable local random variable that represents a hidden factor. It is easy to sample and add levels and depths from these models. You can also use the edge probability for scoring and model selection. However, it is difficult to decide which implicit variable to associate with an input. The following figure shows the different implicit variable generation models [image from Shakir Mohamed's presentation]:
1.3 Application of the Generative model

We need to generate the (generative models) model so that we can move from the associated input to the output for semi-supervised classification (semi-supervised classification), Data manipulation (semi-supervised Classification), fill in the blanks (filling in the blank), image Repair (inpainting), de-noising (denoising), one-shot generate [Rezende et al, ICML 2016], and many more applications. The following figure shows the progress of the Generative model (note that the longitudinal axis should be a negative logarithm probability) [picture from Shakir Mohamed's presentation]:

According to SCALEDML conference Iiya Sutskever's Lecture report "Recent progress in generating models", the resulting model mainly has the following functions: structured prediction, structured predictions (e.g., output text); Much more robust Prediction, more robust predictive Anomaly detection, anomaly detection model-based RL, model-based reinforcement learning

     generates a field that the model can be applied to in the future: really good feature learning, very good feature learning exploration in RL, a quest for reinforcement in learning inverse RL, Inverse enhancement Learning Good dialog that actually works, the real use of the dialogue "Understanding the World", "Understanding the Earth" Transfer learn ING, migration learning