Generate Adversarial Network

Generator & Discriminator

CNN （convolutional neural network，卷积神经网络）

Convolutional networks are simply neural networks that use convolution in place of general matrix multiplication in at least one of their layers.

CNN is one of the deep neural network, which use the convolutional layer to get the information of the picture (sample), input the picture, output the information it got, such as the color, shape, stripe and so on.
So people design one model for generate picture, use the similar method with CNN but with an opposite direction, which is Deep convolutional NN for GAN（DCGAN）. We use DCGAN as the generator, input a random noise, for example, Gaussian noise, and then magnify input to two-dimensional data, based on which then generate the picture.

Use the model of Deep convolutional neural network for GAN (DCGAN) as the generator, add the noise to the original sample to generate one new picture.
Use the model of convolutional neural network (CNN) as the discriminator, get the information of the generated and original picture respectively, and compare the difference between the generated and real picture. Then provide feedback to Generator.

Fig.1 the architecture of CNN used in the cGAN Generator (by Amir Barati Farimani)

If the training data is an image (whose data represents as a 2-D matrix). The encode process is actually the process to find the feature of image using convolutional computation, every time to encode the image, do convolutional computation. After the first encode step (convolutional computation), the 2-D matrix with 64 rows and 64 columns numbers was transfer to 64 matrices and every matrix has 32 rows and 32 columns. After 6 times convolutional computation, the 64*64 matrix was transferred to 512 1-D matrices. The process of decode can be regarded as the opposite direction process of the encode process. In the convolutional computation process, it needs different weight to do the computation. The weight will be adjusted when the objective function computation (including both discriminator and loss function). At the rightest images showed in the figure, both the input row and output row have 512 1-D matrices, they could be regarded as 2 column matrices who has 512 elements. Between the 2 column matrices the full connected networks could be built. There are also some weight should be adjusted when the Full Connected Networks was built, and the adjusting direction was determined by the discriminator and loss function. In the process of decode, except the matrices delivered from the right one, the upper matrices was also added into the below one.