What is the disadvantage of 'self learning by AI' which was also used in completely invincible 'AlphaGo Zero'?

Google's affiliateDeep MindOf GoAI"AlphaGo"Is the one of the top players in 2016, KoreaLee Se-dul (Lee Seong-joo)Mr., in 2017 China'sKiyoshiIt is known that Mr. Win won the victory. In addition DeepMind's new Go AI "AlphaGo ZeroIs pastGame recordWithout giving any data, you will have overwhelming victory of 100 wins and unbeaten with 100 stations against "AlphaGo" with only 3 days of self-learning. For this reason, it is said that "self learning" by artificial intelligence is very versatile and can be applied to anything without relying on past data, but there are also fields that are not really good in practice. I deal with science-related news about where the disadvantages of AI's "self learning" which seems to be the strongest areQuanta MagazineIt reports.

Why Self-Taught Artificial Intelligence Has Trouble With the Real World | Quanta Magazine

"AlphaGo Zero" learned only the rules, learning with only the self-game, without using the data of the opponent actually done by humans. Of course, at the beginning, I start with pointing meaninglessly, but the content I pointed out is "to be tied to victory" and "what does not tie" as new knowledge and eventually Mr. Lee Se-dul, Ke I gained the overwhelming strength of winning 100 with 100 games with "AlphaGo" who defeated Ki.

In addition, DeepMind can learn board games other than Go in December 2017 "AlphaZero"AI has obtained top-level power in chess and shogi only by" self-study ". According to DeepMind, we are not releasing AI software such as "AlphaZero" for the purpose of controlling the world with board games,Room temperature superconductivityRealization "or"Protein folding problemI am doing it as a part of step-up to apply these AIs to "

It is often thought that AI's "self learning" can be applied to anything, but weaknesses also exist. In board games such as Go and Shogi, you can always see the opponent's pieces and the contents that the opponent points to are limited, so it is easy to predict which piece will move. However, AI can not deal with it, as the rules change immediately before the game starts, when "a problem with a high degree of uncertainty" such as "draw a lot for each hand and something happens depending on the content" occurs, the problem still remains It is said to be there.


For example, considering automated driving of automobiles, there are cases where "bad weather such as sudden heavy rain, lightning strikes, storms", "sudden turning direction of a person riding a bicycle on a driving lane", " It will be difficult to judge if AI will respond to innumerable emergency situations, such as "things pop out". There is no problem if AI can actually learn by experiencing themselves, but there are doubts that if you are told that you can experience all the uncertainties that exist in myriad of levels at the level of driving test.


Also, it is expected that patients suddenly have their symptoms worsened during surgical operations, and even if sudden events occur, it is necessary to deal with the symptoms appropriately and deal with it. In these cases, it is necessary for AI to actually conduct a surgical operation and learn a variety of situations, but for the future innovation, it is surgical to AI who is in a suspicious state whether it is insufficiently learned and can respond properly There should not be anyone who wants to provide you with data.

University of MontrealDeep learningIt is a pioneer ofJoshua Benjio"There is a big difference in the" perfect model of the real world "and the" guess-based model "obtained by learning," he says. In response to the numerous "uncertain elements" It is difficult to obtain all the information with "self-study", and many issues still remain.


However, the thought pattern obtained from self learning by AI is obviously different from that of human beings, and it has the potential to create "new ideas" and "concepts that have not been known until now". For this reason, AI's "self learning" is expected to solve the problem which has been regarded as a difficult problem until now, and future development is desired.

in Software,   Science, Posted by darkhorse_log