Artificial intelligence has learned how to create another artificial intelligence

Machines will be able to replace people in the process of creating their own "fellow".

The development process of artificial intelligence has caused many people to worry about the fate of themselves as well as of humanity: they think that we are standing in front of a potential tyrant who holds a strong or gentle power. more, future human work will be dominated by automated machines.

Now, along with the technological breakthrough, synonymous with bad news for those concerned, top researchers are investing in ways to allow software to create a part. Other soft-learning machines. They are on the way to find a revolutionary AI industry software.

Scientists are on the way to find a revolutionary AI software.

In an experiment, scientists at the artificial intelligence team Google Brain used a software to design a machine-learning system for the purpose of testing the benchmark capability of a language processing system. another language. The results show that the new software outperforms the old software designed by humans.

In the last few months, several research groups have also provided information about the process of "creating software to create other software" . These groups include members from the OpenAI nonprofit research organization (co-founded by Elon Musk), MIT, University of California and Google's DeepMind research team.

If this way of building AI can be widely applied, we can push the process of making machine-learning software very quickly. Currently, the cost of hiring machine-learning experts is not cheap, if the machine can replace the human part even in the making of their own "fellow" , it is possible that humans will become an extra factor in the AI ​​manufacturing cycle.

The idea of ​​creating a "learning to learn" software is not so new.

Jeff Dean, head of research at Google Brain, said in a statement last weekend that workers at some production stages could be effectively replaced by a software. He said that "automatic learning learning" technology is one of the most promising research projects his team is currently investing.

"Currently, the way to solve problems including expertise, information and calculations , " Dean told the AI ​​Frontiers conference in Santa Clara, California . "What if we could completely eliminate the expertise on machine-learning?"

A series of experiments from Google's DeepMind team suggests that the "learning to learn " method that researchers are using will reduce the huge amount of data that a machine-learning software needs to be able to operate most effectively.

They challenged their software, requiring it to create a learning system to collect all issues related to a primary goal, thereby requiring it to create a new system design and in that design. They have seen the ability to reproduce and select new tasks without having to go through the usual preparation steps.

The idea of ​​creating a "learning to learn" software is not so new, but previous tests often do not yield the desired results: they are not suitable for human designs. However, this is still considered a potential aspect of artificial intelligence development, said Professor Yoshua Bengio of the University of Montreal in studying this idea in 1990.

AI industry will soon change.

Mr. Bengio said that computer systems are getting stronger and stronger, and with the technology called deep learning, a system of "learning to learn" will have the potential to rise strongly. But he added that such a system would need a tremendous computing power, to replace human experts in the field.

Researchers at Google Brain have also described a powerful system. They use 800 powerful graphics processors to power the software, thereby creating a parity (and even surpass) image recognition system designed by humans.

Dr. Otkrist Gupta, an MIT Media Lab researcher, believes that AI manufacturing will soon change. He and his MIT colleagues plan on an open source software where the learning software will design itself so that a deep-learning system can identify images that are strong and accurate on par with a system made up of people.

"Reducing the burden on the role of scientists will be an effective solution," said Gupta. "It can make us more productive, create more efficient system templates and give us free time to explore ideas at a higher level."