Current Location: Home » news » Technology » Text

"Cat and mouse game" makes the machine more "smart"

放大字体  缩小字体 Release date:2018-04-02  Views:4
Core Tip:At present, the computational power and recognition power of artificial intelligence are rapidly developing, but imagina
At present, the computational power and recognition power of artificial intelligence are rapidly developing, but imagination and creativity are still insufficient. To solve this limitation, scientists have designed a technology similar to the “cat-mouse game” that will make artificial intelligence more “smart” in automatic learning.

 
Cat and mouse game
 
This technology is known as the "adversarial neural network" technology, the United States, "MIT Technology Review" recently rated it in 2018 "one of the world's top ten breakthrough technologies."
 
The recognition ability of artificial intelligence depends on massive sample learning. For example, it can “see” millions of bird images before it can “learn” to identify birds. It is even more difficult to generate realistic bird images. Its limitation is that some things lack mass samples, and such learning also relies on human indoctrination and lack of autonomy. This limits the development of artificial intelligence, especially to higher levels of imagination and creativity.
 
American Ian Goodfellow came up with a design proposal when he was a Ph.D. candidate at the University of Montreal in Canada in 2014. He used two neural networks to conduct a digital version of the "cat and mouse game" - one responsible for "counterfeiting" and one responsible for "Verification," and thus continue to increase in the confrontation.
 
The neural network responsible for “counterfeiting” is called “generating network”. It generates new pictures based on the “seen” pictures, which requires it to sum up the rules, to exert imagination and creativity; and the neural network responsible for “fidelity” is called To "discriminate the network", it needs to rely on training accumulated "experience" to determine whether a picture is a real thing or to generate a network of "self-made" "fake."
 
Generating a network is not "smart" enough at first, for example, it may "think" that birds have three legs, and such "fake" items can easily be discovered. However, with the deepening of machine learning and repeated confrontation exercises, the generation network has a deeper understanding of things and eventually generates works that are “realistic”.
 
Such a neural network has wide application value. For example, in the field of autonomous driving, if artificial intelligence creates mass synthetic images that are close to reality, including pedestrians, obstacles, and other conditions in various situations, the use of these images by self-driving systems for self-training will greatly improve the applicability.
 
Hong Hongsheng, a professor at the Chinese University of Hong Kong, believes that in addition to its specific applications in machine translation, face recognition, and information retrieval, the value and significance of an adversarial neural network lies in its adversarial ideas, which helps improve existing labor. Intelligent Algorithm.
 
From a technical point of view, confrontational neural networks are nearing maturity. Researchers from U.S. chip company Nvidia have trained a system using celebrity photos to generate hundreds of face pictures that do not exist at all but look real. There is also a research team that allows the system to generate van Gogh paintings that look very realistic.
 
While showing great potential, the negative impact of this technology development cannot be ignored. For example, lawbreakers may use such systems to create pictures or even videos to confuse audiovisualization and bring new challenges to supervision and security. Goodfellow said that his current research focus is on avoiding the abuse of this type of technology, and hopes that it will “not go astray”.
 
Liu Chenglin, deputy director of the Institute of Automation at the Chinese Academy of Sciences, said that Chinese research institutions are currently working on the further improvement and optimization of the theory of antagonistic neural networks. There is still much room for development in the theoretical basis, algorithms, and applications of antagonistic neural networks.
 
The Chinese business community is more inclined to apply technology in services and has achieved industry leadership in some areas. For example, use this technology to build a speech recognition framework, or use this technology to generate training data sets to optimize the license plate's precise recognition capabilities.
 
 
[ newsSearch ]  [ Add to Favorites ]  [ Tell a friend ]  [ Print ]  [ Close the window ]

 
Total0bar [View All]  Related Comments

 
Recommended Graphic
Recommendnews
Click Ranking