Home > News content

Facebook off "out of control" AI project: the development of human beings can not understand the language

via:博客园     time:2017/7/31 17:00:32     readed:1462


Editor's Note: This article from the WeChat public number "new intellectual element" (ID: AI_era), compiled Ferguson

According to the digitaljournal reported that Facebook last month caused widespread controversy with the development of human beings can not understand the communication of intelligent dialogue robot project has been closed.

Last month, the Facebook Institute of Artificial Intelligence (FAIR) used machine learning methods to iterate and upgrade the two chat robots, and found that they developed a unique language that humans could not understand. Researchers had to undergo human intervention.

This study once released, quickly in the network and the media has aroused widespread concern and spread, many people exclaimed: the machine to create their own language, human beings can not understand, scrutiny fear.

But some people have pointed out that this is not a mess

Now, this system is locked in the small black house & rdquo ;. According to foreign media reports, now Facebook's research institute has stopped the study of this project, because "worry about the AI ​​may be out of control".

This is not the first time that AI has been deviated from the English language used in training and has turned to developing new and more effective languages. Although in human view, these communication is more like nonsense. However, from the perspective of AI agents, they are actually meaningful.

Communicate in new language: lack of English & ldquo; feedback & rdquo;


The purpose of this research report is that robots can be good negotiators, and they even pretend to be interested in something that is of no value and then sacrifice it in the negotiation process as if he had made concessions. And on the dialogue agent developed from the part of my language, in the research report is really just a half claws, but it has aroused everyone's attention.


As reported by Fast Co. Design, researchers note that chat robots (bots) in the system have completely abandoned their communication in English, but the upgraded version of the system does communicate with other agents. Together they can decide how to proceed with any work. At first, the vocabulary they used seemed incomprehensible, but, through long-term observations, the researchers found that the vocabulary actually represented the task of the agent.

For example, in an exchange, two robots who talked on - Bob and Alice, used their own language to complete the communication. Bob began to say: I can can I I everything else else & rdquo ;. Alice replied, & ldquo; Balls have zero to me to me to me to me to me to me to me to me to me to me "The next conversation is similar to this sentence.

Although the combination of these words does not seem to be practical at first glance, the researchers argue that they reveal how the two robots work and how many times each word they need to do. Bob's repetition explains how he used this term to think about Alice passing more information. If you use this method to elaborate, then it is so much like a lot of people say "I & rsquo; ll have three and you have everything else & rdquo ;.

It seems that agents seem to find that English sentences are not necessary for a particular scene. At present, the AI, which focuses on the depth of learning, follows the "reward" principle that the agent gains the benefit by acting in accordance with a system-specific behavior. However, in this scenario, the continued use of English did not bring the appropriate feedback, so the agent decided to adapt to a more efficient way to communicate.

Facebook AI Research (FAIR) researcher on FastCo. Designs said, "English language, and there is no so-called feedback allows agents to comply." & Rdquo;

FastCo. Designs reported that: "Smart will deviate from their own human understanding of the language, to create their own & lsquo; code language & rsquo ;. For example, if you mention & lsquo; the & rsquo; 5 times, then I want to copy this term 5 times. This is not the difference between the abbreviation of human creation and the abbreviation of human creation. & Rdquo;

Other AI developers also noted that the smart experience using the "abbreviation & quot; to simplify communication. At Open AI, the researchers succeeded in getting AI robots to learn their own language.

Where is the problem?

If AI continues to create its own language, the developer may be having problems creating and using new neural networks, but it is not clear whether this will actually allow the machine to overturn its developers.

However, these new developments allow AI to work more efficiently and if they learn the new AI-created & quot; abbreviation & quot ;, and keep the latest new communication mode, the research team can benefit from the long term.

China IT News APP

Download China IT News APP

Please rate this news

The average score will be displayed after you score.

Post comment

Do not see clearly? Click for a new code.

User comments