Author: Liu Dongyu
link: https: //www.zhihu.com/question/314119766/answer/626691968
Source: know almost
copyrighted by the author. For commercial reprints, please contact the author for authorization, and for non-commercial reprints, please indicate the source.
Deep learning (hereinafter referred to as DL), in other words, can be said to be machine learning based on artificial neural networks. Different from traditional machine learning, DL requires more samples, in exchange for less manual labeling and higher accuracy. DL uses depth to replace breadth, further reduces parameters, and improves fitting capabilities. In many cases, it performs better than traditional machine learning. But DL is the same as traditional machine learning. DL learns a mapping f(x)=y. For example, if x is an input handwritten digital picture, then y is one of 0-9. Specifically, what is a neural network? Arctic Goose: What is an artificial neural network (ANN) zhuanlan.zhihu.com The traditional BP neural network generally refers to a three-layer fully connected neural network, and more than three layers become a DNN (deep neural network). In fact, DNN can solve some problems, but because of too many parameters, it is gradually replaced by other network models: CNN (Convolutional Neural Network), RNN (Recurrent Neural Network). The most successful implementations of the two are ResNet and LSTM respectively. 
The most classic convolutional neural network-LeNet 
cyclic neural network model The following takes the computer vision-related research field as an example, all x is a number of input pictures. If f(x)=y, y is an n-color image, which represents n targets and their positions in x? This is the famous semantic segmentation: the 
schematic diagram of semantic segmentation only needs to modify the CNN structure slightly, first reduce the dimensionality and then increase the dimensionality to complete the mapping. So if f(x)=y, y is a frame, and the position of the bird in the input sky image x is framed? Of course, CNN and sliding window can be done, but it is too slow. So Faster-RCNN was born, which can quickly extract the position of the target in the picture and draw the border.
Use deep learning to draw the image frame. But if f(x)=y, what if x and y are both pictures? In other words, I hope to generate some similar pictures through the study of pictures. What should I do? GAN (Generative Adversarial Network) was born. The idea of this network is to use a network model to generate the target, called a generator; and then use another or more network models to achieve discrimination, called a discriminator. The discriminator improves its ability to discriminate and distinguishes the generated false target from the real target; while the generator improves its generation ability to make the generated false target real, trying to fool the discriminator. The most successful implementation so far is style-GAN. 
Generative confrontation network, say GAN, let's GAN!