A few days ago, the MIT Science and Technology Review published the top ten breakthrough technologies in 2018, and the "Adversarial Neural Network" (GAN) was listed.
What is a confrontational neural network? Why can it be selected into MIT's top ten breakthrough technologies? How is its development context? What is the difference between our previously familiar neural networks? What scenarios can be applied to artificial intelligence? What are the key issues to be addressed? capture?
The adversarial neural network is selected into the top ten breakthrough technologies of MIT2018
Professor Zhang Junping, deputy director of the Mixed Intelligence Committee of the China Society of Automation, member of the Standing Committee of the Machine Learning Committee of the Chinese Artificial Intelligence Society, and doctoral supervisor of Fudan University, gave an in-depth explanation in an interview with the reporter.
GAN phantom in the storyZhang Junping told reporters that although GAN is a "new upstart" in the field of science and technology, in fact, the shadow of this idea can be seen in Chinese and foreign novels long ago.
The idea can be traced back to the novel "The Story of Chess" written by the Austrian novelist Stephen Zweig in 1941.
In the novel, Dr. B, the protagonist, was imprisoned for a long time in a Nazi concentration camp. After trying to get rid of all kinds of ways to get rid of emptiness and loneliness, Dr. B unexpectedly got a chess game.
After he finished the thousands of chess in the book, he used the bread he sent to make a pair of chess and began to play chess with himself. He eventually evolved into a panic of mutual challenge, which made him rise. After he was released from prison, on a yacht, he easily defeated the world champion at the time.
There is a similar shadow in Chinese novels. It appears in Jin Yong’s 1957 martial arts novel The Legend of the Condor Heroes.
Wang Chongtong, the younger brother of Wang Chongyang, was trapped in the cave of Taohua Island by the Dongxie “Yellow Pharmacistâ€. In order to pass the time, Zhou Botong used his left hand and his right hand to fight and entertain himself. The martial art decision is to first "draw the circle with the left hand and draw the square with the right hand", and use the distraction to ensure that both martial arts can be used at the same time, thus making the force multiply.
The key to the protagonist's skill increase in these two novels is "to fight against yourself, trying to beat the other side with all your strength", and the result is that after training, you can easily kill the opponent when you go to the master. Explain in a common saying, "Double fist is difficult to attack four hands."
Let the machine learn to "bet and beat"The principle of the GAN network is essentially the artificial intelligence or machine learning version of the protagonist's practice in these two novels.
There are two roles in a network. In the process of cultivation, the left hand plays the attacking party, that is, the generator, trying to generate a target that is sufficiently similar to the task to be completed in the natural world; the right hand plays the defender, the discriminator, Try to distinguish this fake, generated target from the real target. After repeated hands and hands, the skill of the left hand and the right hand will double, thus achieving the goal of "who is who I am."
Understand this truth, it is not difficult to understand why the GAN network has a feeling of being alone and losing.
It is precisely because the mechanism of the GAN network is "hands and hands, dual-use", so although the initial application scenario is for image-related tasks, the mechanism is universal. As long as you can use this "诀çª" place, you can upgrade your skills to a higher level. However, it should be noted that GAN only uses the hands and hands to train their "hands" skills. In most practical applications, it only uses its own generators, and it has very good results.
Thus, since 2014, the GAN network has been proposed by Ian J. Goodfellow and others, and the various versions that have been deducted so far have spread in various fields like the attacking of the city.
In this year's IJCAI, the top-level meeting of artificial intelligence, ICML and NIPS, and the well-known conference ICLR, the title of the paper can be found in a large number of GAN networks. Image processing, computer vision, natural language processing, speech recognition, smart driving, security monitoring... It seems that GAN is omnipotent.
Zhang Junping uses several examples of artificial artificial intelligence applications.
In the age estimation, GAN can achieve the appearance of human aging or youth based on the given face image through the attack and defense.
In the field of multi-view face recognition and cross-view gait recognition, a similar mechanism is adopted to achieve automatic rotation of the face angle and gait angle, thereby effectively improving the accuracy of multi-view, cross-view face and gait recognition.
In the field of automatic driving, virtual training in complex environments is required for smart cars. At this point, GAN can be used to achieve image generation consistent with the actual traffic scene distribution. Specifically, a random noise image can be input to the GAN first, and the generator is used to minimize the image close to the real scene, and the discriminator maximizes the difference between the generated scene and the real scene. After repeated iterations of the offensive and defensive game, the traffic scene is consistent with the real environment.
"Not only has a breakthrough in the field of application, but also a lot of changes in the way of fighting. Since you can fight with your hands, you can of course fight with three hands or even more hands. You can also group and fight each other. Beat, and so on. You can also swap the circle and the square into other things or so-called functions or structures to fight against each other. But the change is not the same, the internal mechanism is unchanged." Zhang Junping added.
What is the "soft rib" of GAN?"There is no doubt that this technology that can double the 'power' is selected for the MIT's top ten breakthrough technologies. But it is worth noting that this technology still has room for improvement." As the IEEE Intelligent System (Intelligent Systems) and the editors of famous journals such as IEEE TransacTIons on Intelligent TransportaTIon Systems (Intelligent Transportation Systems), Zhang Junping knows the "soft ribs" of GAN.
First of all, it is easy to have problems with bad confrontation. For example, Dr. B. After seeing his strong anxiety and eagerness, the world champion of Chinese chess consciously slowed down the speed of playing chess. The result induced Dr. B's schizophrenia, which caused him to fall into a frantic self-game again, and finally realized his goodbye and bid farewell to the game. The same is true of the confrontation network, and stability has always been one of its problems. Although Wasserstein GAN (WGAN) can partially solve its convergence problem in theory, the actual effect has not reached the point of satisfactory.
Secondly, the guns are the first birds. Since this technology is so bullish, there are many challenges. According to incomplete reports, it seems that fifteen "Wulin Masters" have challenged the GAN network's ability to fight. In extreme cases, adding a pixel to the image may cause the GAN network to misjudge.
Third, the GAN network is also a kind of deep network. On the road of interpretability, it still has not found a very clear direction.
Finally, the basis of the hands and hands is still the hand. And this basic structure has not been separated from the framework of artificial intelligence development in recent decades.
Therefore, Zhang Junping reminded that it is expected that the strong artificial intelligence of "can really reason and solve problems, and that has consciousness and self-awareness" from the weak artificial intelligence will be in the foreseeable future.
Water-cooled capacitor is supercapacitor is a capacitor with a capacity of thousands of farads.According to the principle of capacitor, capacitance depends on the distance between the electrode and electrode surface area, in order to get such a large capacitance, as far as possible to narrow the distance between the super capacitor electrode, electrode surface area increased, therefore, through the theory of electric double layer and porous activated carbon electrode.
Water-Cooled Capacitor,Water-Cooled Power Capacitor,Water-Cooled Electric Heat Capacitor,Water-Cooled Electric Heating Capacitor
YANGZHOU POSITIONING TECH CO., LTD. , https://www.cnpositioning.com