Top of page
Content Area

Moving Forward Together As Your Value Partner

Yoshiharu Habu and NTT's Senior Executive VP - "The Next Move in Developing AI"

Main Content Area

Gendai Business Presents
"The Present and Future of AI" — An Interview between Yoshiharu Habu and NTT's Senior Executive VP

In the summer of 2017, Yoshiharu Habu, a professional shogi player who had just defended his Kisei title, was in Mitaka, Tokyo. His destination was the NTT Musashino R&D Center, a lab that harnesses NTT's knowledge. Here, many unique AI studies take place that are known only to a few.

This article contains an intellectually stimulating interview on shogi, AI, and humanity's future. Hiromichi Shinohara, NTT Senior Executive VP and the head of the Research Division, serves as the lab tour guide and interviewer.

Part I

"I played shogi for 30 years
but never dreamt this would happen..."

1-1. Computers and "onkochishin"

Hiromichi Shinohara Congratulations on defending your title.

Yoshiharu Habu Thank you.

Shinohara Computers relate quite closely to the world of shogi, having improved their shogi skills tremendously. As a shogi player, what do you think about this?

Yoshiharu Habu There are a few reasons why shogi AI developed rapidly. First, the results of development have been opened up to others. Developers engage in AI development simply because they like doing it. They do not do it for commercial purposes; they want many people to use their programs. Thus, open source programs become available one after another, promoting rapid improvement.
Second, people who had no previous relationship to shogi are now developing shogi AI. From chemistry to law, highly skilled people from various fields now participate in shogi AI development. This has led to unprecedented collaboration; we are in an interesting time.

Shinohara How have improvements in shogi programs changed the game for human players? My personal view is that matches between a human player and an AI give the human player an opportunity to discover new moves.

Yoshiharu Habu Yes, our view towards shogi software is changing gradually. At the same time, no matter how advanced AI becomes, it is theoretically impossible to figure out how shogi works completely. Therefore, we are not too worried about AI.
What I have found quite interesting lately is that shogi software is presenting more and more cases that remind us of the word "onkochishin" (from the Analects: to study past knowledge to gain new insights) instead of producing completely new strategies.
More specifically, strategies that were frequently used 100 or 150 years ago but that were considered absolutely useless recently are now increasingly being re-evaluated and explored to develop their potential.

Shinohara So, do these re-evaluations of strategies lead you in new directions? For example, do the strategies being re-evaluated and the original strategies suggest different plays five or ten moves later?

Yoshiharu Habu I would say they are slightly different; they differ a bit from the original. When we try the re-evaluated versions, they tend to make sense. This suggests that the advancement of shogi programming may result in a reevaluation of the past 400 years of shogi plays.
Of course, shogi software sometimes makes unprecedented, innovative moves as well as moves that contradict the theory of shogi, but we usually reject these moves based on our human aesthetic sensibility.

Shinohara Am I correct that by aesthetic sensibility you mean the sensibility of shogi players to dismiss certain moves because they are unartistic and to dislike certain ways of arriving at checkmate? I do not think such unconventional moves will disappear no matter how advanced AI becomes in the future. What is your opinion?

Yoshiharu Habu I agree. However, exposure to shogi software may impact our aesthetic sensibility.
Aesthetic sensibility never stays the same; it continues to change over time. Therefore, it changes gradually. In fact, some piece placements that were previously considered "ugly" or "bad" actually work quite well at times.

1-2. Determining an acceptable range of variation

Shinohara What is your relationship with shogi software?

Yoshiharu Habu I hardly use it. To be honest, I still have no idea how to use it or how much I should use it.

Shinohara I see. Do good moves come to your mind more frequently as shogi software becomes better?

Yoshiharu Habu Better shogi software definitely broadens our range of ideas. Computers show us moves that we would never even think of-- sometimes they make moves we would never hit upon even after thinking for 100 years.
The issue here is that it is difficult to decide how good or bad a certain match state is. It is of course difficult for us humans to decide, but the same applies to AI. Sometimes a certain match state looks advantageous to a computer but I perceive it very differently.

Shinohara This is not about the aesthetic issue from earlier, right?

Yoshiharu Habu No, it is not. The moves indicated by computers are not necessarily "the" answer. A problem called the horizon effect describes how AI's evaluation accuracy drops once it reaches a certain point. So, we do not know how much we should trust the evaluation results output by AI.
Since the values output as evaluations of match states are not absolute, I feel that current shogi players must determine an acceptable range of variation for such evaluation results.
I have played shogi for 30 years, but I never dreamt that we would live in a time in which we would need to think about such variation (laughs). A long time ago, we had no databases or the Internet; we studied shogi by copying match records and moving pieces accordingly.

Shinohara We live in demanding times (laughs).
AI lays out many possibilities, but it is us, the humans, who make choices. You just mentioned that AI will change human aesthetic sensibility gradually, but this does not mean that each of us will arrive at an identical aesthetic sensibility. Each individual will still be unique.

Yoshiharu Habu That's right. After all, AI is producing statistically better results but not absolutely correct results. This is an extremely important point.
When AI becomes part of our daily lives, we will likely mistakenly believe that answers produced by AI are correct. Even things that would be criticized as incorrect if suggested by humans may be considered correct if they come from AI. It will be challenging to figure out how to introduce AI to society while ensuring such misconceptions do not occur.
When we think about the process of evolution, standardization aimed at deciding that a certain thing is correct is clearly the wrong path. How to ensure diversity is a quite important matter.

Shinohara When people make decisions, their decisions are often influenced by their emotions or ego, which are human characteristics that do not present issues of right or wrong.
As you recommended, technology should help multi-faceted people act in multi-faceted ways by promoting human uniqueness while making the most of AI.

1-3. Can the mechanism of "inspiration" be explained?

Yoshiharu Habu As shogi software becomes better, it is starting to show like a mirror how our minds work. Comparing ourselves with computers gives us insight into aspects in which people are better than computers as well as where our blind spots are.
AI development is modeled based on the human brain as seen through neural networks. So, if technology continues to advance, will a day come when our difficult-to-verbalize ability to find inspiration is explained? Or, will it remain somewhat mysterious, as we now believe?

Shinohara At present, AI does not produce answers based on a true understanding or interpretation of a given subject. To me, AI's limit is that it always produces an answer, while people can say "I am not sure" or "I do not know" when they do not have an answer.

Yoshiharu Habu In other words, AI can think correctly based on flawed data and produce incorrect answers.

Shinohara That's right. For this reason, I think AI is far from achieving semantic understanding.

Yoshiharu Habu I often think that the concept of time differs completely between humans and AI. Humans sleep more or less one-third of the day, but AI can keep learning without breaks. AI can even learn at super high speeds. Given this, AI may make the impossible possible if provided with the right framework.
On the other hand, using the example of shogi to think about the characteristics of human minds, becoming better shogi players means reducing the number of moves we consider.
Even though there are a hundred possible moves, shogi players become able to pick from among two or three of them instantaneously. Skipping unnecessary steps is part of getting better.

Shinohara I agree. Some parts of the brain think logically, while others take leaps. When people engage in activities, as you said, not everything proceeds logically. We make decisions that are rather far-fetched. This is humanity's advantage and I think it will remain so for quite awhile.
Now, please allow me to show you some of our research activities.

Part II

What is the "global Prince Shōtoku" that Yoshiharu Habu encountered at NTT's lab?

2-1. Experiencing AI technology's expansion

Yoshiharu Habu, a professional shogi player, visited the NTT lab in Mitaka, Tokyo. There, he was introduced to a wide variety of unique AI research activities. Following Part I, which covered shogi and AI, in Part II, Yoshiharu Habu and Hiromichi Shinohara, NTT Senior Executive VP and the head of the Research Division, exchange opinions about current AI research and how we should live with this technology.

2-2. Is AI humanity's enemy?

Yoshiharu Habu After taking in the whole range of research achievements, I truly feel that AI research has advanced in many fields. In addition, I suspect that these achievements will come into society in practical form very soon.
When it comes to AI, we tend to have an image of "humans versus AI." In shogi, it becomes big news whenever shogi software defeats a professional shogi player. There have also been arguments that AI will take jobs from humans. However, that is not the whole story, is it?

Hiromichi Shinohara As you said, "humans versus AI" is one way to look at it. However, our research is based on the notion that AI exists to support us, and it reinforces and brings out our abilities.
Needless to say, computers exceed humans in terms of computational ability and memory. Our objective is to effectively tap these capabilities to create an AI that coexists and co-creates with humans.
"AI" is just a single term, but this technology has expanded into numerous fields. At NTT, we have conducted various types of AI research; we collectively refer to them as "corevo." Corevo consists of four types of AI, each making the most of NTT's strengths as a telecommunications carrier.
The first type of AI is "Agent-AI," which works as an agent for humans. It applies voice recognition and linguistic analysis technologies to human-provided information in order to support people.
Last year, there was an automated voice recognition challenge in the US in which participants competed to perform human voice recognition despite noise. Our group won the competition with a wide lead over our competitors. Since we were originally a phone company, we naturally have extensive knowledge about sound. A specific example of an application of our voice recognition technology is something to assist call center operators.

Yoshiharu Habu The technology I just saw truly amazed me. It recognized all types of speech made simultaneously in different languages. While I only heard a large amount of noise, the AI recognized speech in Japanese, English, Chinese, and so on. That really was a "global Prince Shōtoku" (a figure in Japanese history who could purportedly carry on conversations with 10 people at once) (laughs).
No matter how hard we try, we cannot improve our hearing or eyesight any further. Enhancing sensor technology will enable us to recognize what we could not before.

Shinohara I believe that Agent-AI will be put into practical use very soon. However, we should note that we have inner, unspoken desires in addition to information such as words that we physically output. For example, even though someone says "Yes," you may realize that the person actually thinks "No" when you look into his or her eyes.

Yoshiharu Habu Yes, that can happen.

Shinohara It will take more time before Agent-AI can respond to such deeper feelings.

Yoshiharu Habu I tried the AI daily conversation technology. This is a typical example of something that is easy for humans but difficult for machines. However, I thought the AI that I tried spoke quite naturally.

Shinohara The AI I introduced today can carry on conversations at a kindergarten level. It is still too difficult to have it say something witty or that makes people happy. It is also far from able to speak appropriately based on the time, place, or occasion. Still, it may be helpful in some way if, for example, it could respond when spoken to by elderly people who live alone.

Yoshiharu Habu That kind of AI can add color to daily life. Today's aging society has many people living alone. It would be great if this kind of research continues.

2-3. How to improve AI's reliability

Shinohara The second AI we propose is "Ambient-AI." Ambient refers to our surroundings. We are developing an AI that fully understands people, things, and the environment in order to instantaneously predict and control the short-term future.
Our research includes an AI that automatically detects dangerous driving, an AI that sounds an alarm upon detecting signs of equipment failures, and an AI that predicts product quality in manufacturing processes. Recently, AI-assisted taxis have become news. Such AI shows taxi drivers the routes where they are likely to find customers.

Yoshiharu Habu I saw it on NHK the other day. So that was NTT's technology!

Shinohara This AI uses mobile phone signals to plot the distribution of people. Such distributions used to be processed as data. Now, this data can be used to generate something like a heat map, which is then processed as an image. This enables high-speed calculation. A taxi company in Nagoya uses this technology to produce good business results.

Yoshiharu Habu Improvements in computers' operating capabilities are something we can easily understand. But at the same time, data compression technology, such as that for processing data as images, is also advancing even though we do not see it.
The AI I witnessed a few minutes ago was also interesting. It searches for a supernova by examining images taken by the Subaru Telescope. AI can detect characteristics in the blink of an eye that we cannot recognize on our own.

Shinohara That's right. One matter of concern is that AI may pop out a result without explaining the reason for it. We may find it difficult to simply accept such answers. Therefore, we must consider how to explain these answers to people so that they feel satisfied.

Yoshiharu Habu I understand your point. In the field of medicine, a single mistake may lead to death. Though AI can pop out results, I think how people perceive such results is a difficult challenge.

Shinohara There are various types of AI technologies, including deep learning and neural networks. The most important issue here is how to prepare perfectly correct data. Incorrect data compromises the reliability of every answer.

2-4. Exploring the human mystery

Shinohara The third type of AI is "Heart-Touching-AI."
This AI analyzes the human mind and body to understand the subconscious and instincts that even its users are unaware of. Then, the AI provides feedback to users to help them do things they could not previously.
More specifically, this AI is for brain science in sports. The brain is what separates elite athletes from mediocre athletes, although of course muscles and other elements also count. Body movements follow the brain's instructions. Therefore, we are studying how the brain influences sports.

Yoshiharu Habu I tried the system that reads a batter's mind based on the batter's unconscious pupil movements. I did not know that kind of technology existed, but I thought it had strong potential.

Shinohara We can now measure how nervous players are in actual matches. We hope to produce interesting results to answer questions like "What should we do to achieve the best performance when nervous?" and "What tricks improve performance?"

Yoshiharu Habu As the Tokyo Olympic Games will be held, I look forward to hearing the answers.

Shinohara When people think about AI, they usually pay attention to its abilities to think and produce answers. However, before we think and produce answers, there is in fact the important step of recognizing and understanding the world around us.
For example, when you are at a noisy party, you still hear your name if someone says it. We do not know exactly how this process works, but it is not the case that we first understand the sound through the ears and then pay attention to it; rather, there is some kind of filter in place before such understanding occurs.
Heart-Touching-AI exists to study the mystery of human inputs and outputs, including the example I just described.

2-5. Connecting AI with AI

Shinohara The last type is "Network-AI." We use it for two different purposes.
The first purpose is to use AI technology to create stable networks that do not stop.
The second purpose is in fact our true focus. You know that AI can produce one-sided answers, right? If this happens, different AIs should connect with each other and produce answers as a collective intelligence. For example, when extensive damage occurs such as that caused by a mega-quake, the answers produced by an AI in the Sendai area and an AI that takes into account all of Japan naturally…

Yoshiharu Habu …Diverge

Shinohara In such a case, we must reconcile the answers. Our present attempt aims to create a consolidated AI that can produce both overall and locally optimized answers. We are currently working on weather information by exploring how to produce new weather information by skillfully connecting the weather of a vast area with local weather.

Yoshiharu Habu AI technology truly has infinite potential.
So, for example, if we have a huge volume of data that shows us human behavior patterns, would it be possible to predict the number of people in a town and what the town would look like one hour from now?

Shinohara We are thinking about using AI like that. When a major disaster occurs, predicting the short-term future is very important because taking measures optimized for a group of people enables them to move smoothly, while taking measures optimized for individuals results in everyone moving in the same direction, causing massive congestion.
When we think about everyday living, rather than unusual events like disasters, do you think humans find it enjoyable to do similar things because decisions have been made using similar criteria? This is no fun at all.

Yoshiharu Habu Not at all.

Shinohara So, as I explained at the beginning, our objective must be to create happier, more fulfilling lives with AI technology.

Yoshiharu Habu Okay. So what fields do you think offer the most promising services?

Shinohara For instance, when we consider the issue of the aging society, AI may become useful in trying to figure out how to create comfortable, convenient environments for everyone.

2-6. How should AI play a part in society?

Yoshiharu Habu AI is best at optimization. I wonder if AI can clarify what humans cannot or make new discoveries by finding answers under specified conditions or by calculating astronomical figures.

Shinohara Yes, AI can do this because it is best at finding optimal solutions as well as the minimum value from vast volumes of data. One of our urgent tasks is to find fields in which AI can produce socially useful answers by employing this kind of technology.

Yoshiharu Habu We can think of many things, like how to ease traffic or prepare train schedules.

Shinohara We have already started experiments to study how to predict and solve traffic congestion in China and London. The problem is the extent to which humans listen to AI. They may say, "No, I like the other way better" (laughs). Aside from the technical problems, there is a human problem-- people do not listen to AI unless they are directly presented with the benefits.

Yoshiharu Habu If you accumulate data that suggests approximately 20% will not listen to AI, you may be able to proactively respond in suggesting routes.
AI is being developed globally, and I think that AIs reflect the culture of their countries and regions. This may be a stereotypical view, but people in Europe and the US think of AI as something like what is shown in The Terminator, while people in Japan think of it as something like the anime character Doraemon.
My personal opinion is that conflicts between humans and AI will be minimized if we prioritize humans and introduce AI from the perspective of how AI should operate in people's daily lives. I hope that is our future.

Shinohara AI stands for "Artificial Intelligence." When we think of supporting people and improving their lives, there is also IA, which means "Intelligence Amplifier." This reinforces and supplements people's functionality. I will promote our R&D efforts to advance in that direction.

Yoshiharu Habu Today, I visited your lab and really felt the potential of NTT's fundamental research. This was a fascinating experience. Thank you very much.

Shinohara It was a great pleasure to have you. Thank you, and please visit us again.

Watch the interview