Published on Jun. 29, 2017: The Denki Shimbun (The Electric Daily News)
Shojiro Matsuura
President & CEO

These days, I am frequently amazed by news of the younger generation accomplishing achievements to levels previously unimaginable in sports and games. That includes table tennis, gymnastics, ski jumping, bouldering, and such ancient board games as Go and Shogi. I have a feeling that a major change is happening, although it is unlikely the Japanese people experienced any dramatic improvement in their average levels of basic physical or brain functions.

Have there been significant advances in training or learning systems? If training systems rapidly developed and progressed—making it possible to observe people's development and growth in great detail and precision, merge those observation results, and optimize physical function development in step with advances in brain functions—then it is unsurprising that young sports and board game players are performing well.

I recently came across a news article that had stuck with me (it was a front-page feature in the June 4 morning edition of The Nikkei), and it gave a perfect example of what I described above: artificial intelligence (AI) had overwhelmed the world’s best professional Go player over a three-round match.

It is commonly known that research-and-development efforts are underway to realize advanced AI by modeling the human brain’s neuronal system, i.e. the signal transmitting system, and in turn equipping computers with self-learning (deep-learning) functions. Results have already shown that such R&D efforts exceeded levels of the world's leading chess or Shogi players. Meanwhile, it was thought that more years were needed for AI to surpass Go grandmasters, given the various options available in the game; but contrary to such expectations, AI has already become capable of demonstrating its high potential. Apparently, what is behind this progress is advances in deep-learning functions that were achieved through reinforced information processing technology and self-learning functions.

When solving problems, one draws multiple conclusions from various options and accordingly reaches the best possible conclusion. The news of AI winning against a human Go player demonstrated that improved AI learning functions are enabling to quicken the process of deriving conclusions.

In order to speed up this conclusion-drawing process, AI must fully identify an extremely wide range of relevant events. Then, after integrating interconnectable cases into the full system of the subject matter, it must be capable of finalizing an option that not only best meets requirements but also is within the extent of restrictive conditions and the time limit. Today’s computational science and computer technology has apparently reached a level where people can expect advanced AI to turn into a reality.

The achievements of the young athletes I mentioned earlier are not based on computers. Rather, I assume the approach they take requires them to harness advanced observation methods and experience-based knowledge to comprehensively identify—from the activities and behavior related to their field—the combination of actions best suited for each player. Then, he or she presumably practices (i.e. learns) that action to the fullest extent. This can be perceived as one example of deep learning.

Ensuring safety (i.e. reducing risk) in nuclear power operations is the overriding and crucial issue in the nuclear arena at the moment. It goes without saying that this issue is more difficult in identifying details and in dealing with, compared to sports or games. Then again, deep learning essentially has the same significance and effect because sports, games, and risk reduction all can be traced back to the actions and behaviors of people.

Regarding the process followed in the wake of a nuclear safety issue, which ranges from identification and specification of initiating events, the possibility of post-event development, and response for addressing the issue, the events that occur and relevant phenomena and conditions span a broad spectrum. Even so, it should be possible to draw highly plausible conclusions if efforts are made to fully use latest computer technology and evaluate big data (e.g. fundamental knowledge and phenomena already established and acquired in each area) under several probable conditions.

The method of probabilistic risk assessments (PRA) currently underway is expected to serve a critical part in deep learning. Putting AI into general use previously was an impossible dream, but it apparently could happen earlier than expected.

end