In 2014, Ian Goodfellow proposed the concept of generating a confrontation network (GAN). Since then, the generation of confrontation networks has been a research hotspot in academia. Yann LeCun also called it “the most exciting area of ​​machine learning in the past decade.†idea". What is the current development of GAN, what can it do, and what are the future prospects? This article was originally compiled from Hackernoon's article entitled "The New Neural Internet is Coming."
What is GAN - Generated against the development of the network
Neural network is a very popular science and technology hot word recently. Its core purpose is classification. The classifier is a machine that automatically classifies input values. The classifier enters a numerical vector called a feature (vector). The output of the classifier is also a numerical value that represents the result of the classification. The goal of the classifier is to make the proportion of the correct classification as high as possible. The generation confrontation network (GAN) consists of a generation network and a discriminative network, and learns by letting two neural networks play against each other. It is a method of unsupervised learning.
The generation network takes random sampling from the latent space as input, and its output needs to imitate the real sample in the training set as much as possible.
The input of the discrimination network is either the real sample or the output of the generation network. Its purpose is to distinguish the output of the generation network from the real sample as much as possible.
To generate a network, it is necessary to deceive the network as much as possible.
The two networks confront each other and continuously adjust the parameters. The ultimate goal is to make the discriminating network unable to determine whether the output of the generated network is true (until Nash equilibrium is reached).
If we treat a typical neural network (such as an image classifier) ​​as the left hemisphere of the "brain" of a neural network, the right hemisphere that produces the confrontational network is like a brain - a hemisphere that is responsible for creativity.
GAN is the first step in cultivating the "creativity" of neural networks. A typical GAN ​​generates images based on specific keywords from random noise or latent variables. At present, the quality of images generated by GAN is not good and the resolution is limited. But recently NVIDIA has made new progress: it is possible to generate realistic images at high resolutions, and they have opened this technology.
Conditional GAN ​​and Variational Self-Encoder
There are many types of GAN, and the complexity, structure and abbreviation are different. What people are most interested in are conditional GAN ​​and variational auto-encoders. The conditional GAN ​​can not only imitate the "classroom", "face", "dog" and other large classification images, but also produces a more detailed image classification. For example, the Text2Image network can convert textual descriptions of images into images.
GAN generated fake room, fake dog, fake celebrity, and fake art results
By setting a random seed for the "meaning" vector, we can generate an unlimited number of matching bird images.
Enter the text description "This bird has a white chest, a light gray head, black wings and tail". Text is converted into a vector. Add a random seed to complete the GAN training. Give the result
Our immediate opportunities
Imagine the world two years later. A company like NVIDIA will develop GAN technology to a mature industry standard, just like the celebrity avatar we can see now. This means that GAN can generate any image at any time as required. You just need to give a description of the text. So many photography and design related industries will be out of date. Please see the picture below for decomposition.
Given different random seeds, this neural network can generate an infinite number of image results.
On the left is the text description entered, and on the right is the result generated by GAN.
Enter "I need a photo of a cow and dolphin fit." Get a picture.
Enter "I need a picture of cattle and dolphins separate."
Enter "I need cattle and dolphins to separate and enjoy the photos of life." Get the picture three.
Super personalized
What's chilling is that such a network not only enters the description of the target it needs to generate, but it also receives a vector that describes you and the target consumer. This ad can in depth depict your personality, your web browsing history, recent transaction history and geographic location. Therefore, the result of the one-time generation of GAN is specially created for you. The user’s click-through rate will surely burst.
Enter “Girls in our branded clothing†+ blonde, California, dream red Ferrari, priority white, and the result is a map in the lower left corner. Enter "Girls Wearing Our Branded Clothing" + red-brown hair, recently searched for Chanel, and plan to travel to Monaco. The result is a map in the lower right corner.
After “measuring†your reaction, the neural network will adjust to make the advertisement more and more accurate to your taste, stimulating the point that makes you most excited.
Bubble trend
So, one day, there will be completely personalized content everywhere on the Internet.
What everyone sees is a personal adjustment based on his own lifestyle, opinions, and personal history. In the U.S. presidential election, we have witnessed the intensification of this bubble pattern, but the future situation will only get worse. GAN can generate a variety of content for individuals, and is not subject to media restrictions - from simple image ads to machine-generated complex ideas, posts and published works, creating a continuous feedback loop, constantly improving based on human-computer interaction . There will be competition between different GANs - a fully automated war, and we humans are the battlefield. The driving force behind this trend is very simple - profit.
This is not what scare people's doomsday speech. The horn of war has already been heard. It's just that we haven't seen smoke yet.
Is it good or bad?
I do not know. However, there are some things that are in no hurry: the arrival of this kind of technology is inevitable, and extensive public discussion must be conducted as well as preparations for a successful stop. So, it is best for us to begin thinking now - how can we benefit from this process while benefiting from it?
Technical aspects
For some technical limitations, the future described above has not come true. At present, the image quality generated by GAN is still very poor and it is easy to be found to be false. However, NVIDIA has demonstrated to the world that it is feasible to generate 1024x1024 realistic faces. If the technology is to be further developed, we need faster and larger GPUs, more theoretical research on GAN, more ingenious methods of training GAN, more mark data sets...
Note: We do not need new energy, quantum processors (but this may help) or artificial intelligence to help us achieve technological advancement. The resources we need can be completed within a few years, and some large companies may already have these resources.
In addition, we need smarter neural networks. The progress of GAN will be applied first to super-resolution technology, bringing great benefits to the progress of super-resolution technology.
Dc Cross-Flow Fan,Dc Axial Fans,Dc Cross Flow Blower,Dc Cross Flow Fan
Original Electronics Technology (Suzhou) Co., Ltd. , https://www.original-te.com