Self-learning "Like a human" What is the key to developing the next generation artificial intelligence?

ByAngelo Yap

At last the computer wins a professional shogi player with goAs has been talked about recently, remarkable evolution of artificial intelligence technology has become a hot topic in recent years. What you need to evolve such an artificial intelligence an even further,FortuneI am exploring.

Why Memory and Mimicry Are The Next big Frontiers in AI - Fortune
http://fortune.com/2015/12/04/next-ai-frontier/


In the past few years, researchers involved in artificial intelligence development have been trying to make computer and voice recognition possible by computer,Deep learningI have learned various kinds of information on a computer using. However, the image / speech recognition function of the computer has already evolved to a considerable level, and it is clear that these functions will gradually improve in the future. In addition, most of the difficulties in algorithms and fundamental research that we have had so far have been clarified.

ByHyoin min

Particular attention was paid to artificial intelligence related technology in 2015 by Google and Facebook, which are major technology companies, and the two companies are showing new products that will be a big breakthrough, respectively. The innovative technology developed by Facebook is the artificial intelligence algorithm published in the fall of 2015 "Memory NetworksSo, like a baby, it helps computers to learn. On the other hand, Google released a technique that lets computers learn in a way different from Facebook, adding time and memory according to the situation and problem. Google has applied this artificial intelligence technology to its translation tool,Inbox by Gmailof"Smart Reply"It also applies to the function called.

Smart Reply is a function that can be used with Inbox by Gmail to analyze the text of incoming mail and present three kinds of reply corresponding to it.


Other things that attracted attention in artificial intelligence related fields are startup "Osaro" that develops machine learning software for industrial robots. At the end of 2015, Peter Tier, a well-known investor in Silicon Valley, Scott Bannister, a businessman, and Yahoo! Mr. Jerry Yang and others, totaling $ 3.3 million (about 400 million yen)FinancingSucceeded and became a hot topic. Osaro is "Deep · Reinforcement · LearningIt is called deep learning andReinforcement learningIt is a kind ofQ learningTo develop products to evolve from a research level to a product level.

Osaro co-founder Derrick Pridmore said the computer needs a clear policy to clarify what's coming next. Up to now, engineers have crafted this policy cleverly, but it seems that there is no extensibility in these policies. There is "Deep · Reinforcement · Learning" useful there. Deep · Reinforcement · Learning is a learning method to let the computer work perfectly, so that you can learn the correct action and produce the correct result.

Incidentally,DeepMind acquired by GoogleDevelops "DQN" which is the most famous example of Deep · Reinforcement · Learning. DQN makes it possible to hiht high scores by simple games etc. by repeating learning automatically without human's hands. Details of DQN can be confirmed in the following article.

When will the day the artificial intelligence "DQN" that learns the game by himself learns more than humans threaten humans? - GIGAZINE


These kind of researches have not been able to leave the field of research and development so far. Indeed, many technology companies are still unable to escape the research and development area in the development of artificial intelligence. However, Osaro's Mr. Predmour plans to build a production system to provide learning necessary to enable robots to recognize the situation and cause actions based on the information they have recognized, skipping research and development It is.

Software that Osaro can develop to commercialize can theoretically teach "Robots can learn from tasks and tasks". Even better, software can adapt learned knowledge as a parameter for each environment and use it. The key to this technology is that each robot is automatically optimized in the use environment of a different robot by the user in that the parameters can be learned instead of being programmed That's why.

ByPatrick Lentz

DeepMind's DQN will advance learning by randomly playing games, but if you do such things in the real world, it will take a huge amount of time and money. Therefore, Osaro seems to have created an algorithm to support the computer learning process with "imitating humans". Of course, it is not only Osaro that imitates the human learning method, but the University of Washington asks children to learn robotsAnnouncementdoing.

"Minds.ai"The reinforcement learning is at the forefront of the artificial intelligence field, reinforcement learning is to reinforce poor artificial intelligence," said Sumitt Sayyar, CEO of startup called " As it seems, multiplying reinforcement learning and deep learning seems to be the key to the birth of the next generation of artificial intelligence.