The generation of self-encoder against network Aegan

the generation of self-encoder against network Aegan 1 Introduction

Autoencoder is a good special functions extraction model, however, if we use Autoencoder's decoder part as a generator, the effect is not good, the main reason is that we extract the H itself is a heuristic feature, without any explicit meaning, So when we use decoder to do generation, we can't give a meaningful distribution of H.
Gan is the most good generation model, there is also a problem, is the production of images, lack of reasonable interpretation, and is not particularly true.
Therefore, in this article, the author proposed a new Aegan, using Gan form to enhance the encoder generation ability, and the use of autoencoder to achieve image reconstruction, so that the resulting image more real. 2 Aegan

The whole Aegan can be seen as three parts: Encoder,decoder,discriminate. The structure is shown in the figure.

Encoder: Encodes the true image X to obtain a hidden layer Z.
Decoder: Decoding the encoded hidden layer Z to obtain a reconstruction of the original input x x^ \hat{x}
The above two parts make up the autoencoder part of the Aegan, and train it with the traditional method of minimizing the reconstruction error.
Discriminate: A discriminant is adopted to determine whether the input z is from a distribution we generate, or a feature extraction of a real value. Comprehension

Given a D, to determine whether the input z is from a real sample, or from a particular distribution, which is fundamentally different from the original Gan, the original Gan is to fit the distribution of sample pdata P_{data}, and Aegan is to let us extract the feature z near our given distribution of PZ (z) p_ Z (z). This can be a good solution to the original Autoencoder in the generation of the middle layer H features no objection to the problem.
We then use the feature of our constraint to reconstruct it so that we can get a new sample that is infinitely close to the original sample. This is good to eliminate the original Gan generated image is not realistic problem. weakness

There are advantages, there is also a shortage of nature. Although the reconstruction error solves the problem of the authenticity of the generated sample, it also makes the diversity of Gan generated greatly discounted, which is a contradiction in itself.