©You Sung Gil

Professor at Future University Hakodate, Matsubara Hitoshi is a pioneer in artificial intelligence (AI) research. He has been engaged in introducing the potential of AI through familiar themes such as computer-played shogi and Robo-Cup soccer. Recently he stirred surprise with a short-form novel written using AI that passed the first screening for the Hoshi Shinichi Award, a literary prize for short science fiction. Matsubara says that he is aiming at the study of sensitivity.

The Meeting of Atom and Freud

©You Sung Gil

In kindergarten I was entranced by the television anime program "Tetsuwan Atom" (Mighty Atom) about a boy robot. Dr. Tenma, the inventor of Atom, became my ideal. My father told me that Dr. Tenma's profession was "engineer," so from the time I was in kindergarten I always said I would be an engineer when I grew up.

In junior high school I started reading the works of psychologist Sigmund Freud. I got the idea from Kitayama Osamu, a musician and songwriter I liked at the time. I was only a teenager, so I'm not sure how much I understood, but I thought it was really fascinating how much is not understood about the human psyche.

In university, I wanted to link my two interests--in Atom and Freud--together, so I got the idea of making a robot with both intelligence and feelings. To accomplish that I learned that I needed to study the "brain" of a robot, and that meant pursuing the field of artificial intelligence.

Tokyo University's First AI Major

I enrolled in the Faculty of Science at the University of Tokyo (Todai) in 1977, but when I said I wanted to study artificial intelligence, I encountered stiff resistance. One professor went as far as to call AI "rubbish." But in those days, that professor wasn't alone; most people dismissed it entirely. The 1970s, in other words, were the doldrums as far as AI research was concerned.

Computers were originally invented in the 1940s, particularly for the purpose of accurately performing calculations of massive amounts of data for military purposes during World War II. After the war ended, the technology began to pass into the civilian sector. In the United States and Europe in the 1950s and 1960s research experimenting with use of computers for deductive reasoning was quite lively. That was what is known as the "first AI boom."

Moving into the 1970s, however, because the results of previous research did not live up to expectations people began to say that AI didn't offer much hope. So in Japan, which was in the midst of rapid economic growth at the time, there was no AI boom, and the general word was put out that there was no future in AI research.

I was actually something of an amanojaku--a maverick or perverse fellow always going against the tide--determined to pursue some minor path. So the very idea that the subject was one that people had decided to dismiss and disdain caught my fancy. I'm the type--the more people oppose something I want to do, the more attached I get to it.

When I went on to graduate school at Todai, I thought: now I am going to pursue AI research in earnest! I applied to study under an engineering faculty professor famous for his research on robots. I saw that he had included "artificial intelligence" in his list of research themes.

I was accepted to study under the professor, but then he told me, "I don't know a thing about AI, so you're on your own." He explained that he had learned about AI while pursuing research at MIT in the United States, and since it seemed interesting, had just added it to the list of his interests. Still, he said I could do it! That alone made me happy. I'm grateful to this day that he gave me that chance. And so I became the first graduate student majoring in AI at the University of Tokyo.

Japan's First AI Boom

©You Sung Gil

Today, universities have regular courses on the subject of AI, but when I was at Todai, there was nothing. We did have an "independent seminar" on AI, however, at which students studying under various professors interested in the subject would get together. We pursued our interest, gathering weekly on Saturday afternoons in a university classroom. The group had started out with five, but when I joined, there were around twelve. There were hardly any books on AI in Japanese at the time, so we ordered expensive books from overseas on the subject and perused articles that were published in English.

It was not long before development of AI programs involving specialized expertise in medicine, law, finance, and other fields--known as "expert systems"--became very lively in the United States. There were high hopes for the business potential of using computers to perform the services of doctors and lawyers, and in the latter part of the 1980s, this stoked another AI boom worldwide.

That boom came ashore in Japan as well and the number of students in the "independent seminar" on AI burgeoned. Anyone could join, and it soon swelled to 100. It was the second AI boom in the world, but the first boom in Japan.

Out of the Doldrums

In the early 1990s it became clear that "expert systems" would not feasibly replace human expertise, and interest in the subject faded. Computers, unlike human beings, lack ordinary common sense and can therefore jump to alarming conclusions. For example, when set to diagnosing treatment for a person with a fever, the computer prescribed: "Let the patient die." In Japan as well, the boom vanished with the bursting of the Bubble of the overheated economy. Society was not interested in AI and funding for research was zero. The number of AI researchers declined. It was part of the recurring cycle of boom times and doldrums for AI research.

But today, with the advent of the third AI boom, interest in AI is once more lively worldwide, including Japan. The heightened interest was triggered by the publication in 2006 of an article by University of Toronto professors Geoffrey Hinton and his associates proposing the idea of "deep learning." Their findings enabled improved precision in the way computers could "learn" principles from data, even without being given instruction by human programmers. It proved especially effective in image recognition, and in 2012, it resulted in computers being able to recognize images of human faces from image data with far greater accuracy than could humans. This capability was widely acclaimed, reigniting hopes among some that artificial intelligence can be used to solve all the problems in the world.

People who had turned such a cold shoulder on AI research suddenly became friendly and interested. Personally, I think that the "doldrums" may come again, and I worry that cold shoulders will again be turned upon our research, but I am glad that there is an increase in the number of people who have taken an interest in AI. The people who form the middle echelons of the field today are those who were educated during the boom in the 1980s and 1990s.

My fondness for AI research is in part precisely because of its minor status. Still, I'm devoted to the field and eager for it to advance. For that to happen, I want more people to be involved in research and I want AI to be recognized as a major field. I admit to feeling these two contradictory impulses.

専Research Understandable to Non-Specialists

©You Sung Gil

Writing computer programs is the basis for AI. It's quite sober, inconspicuous technical work that ordinary people have a tendency to consider "difficult" and "mystifying." It is important for research to be sober and difficult, but if that is all, the world will never understand what we are trying to do. Research funding won't increase and we will be plunged back into the doldrums.

To prevent that from happening, I have aimed for approaches that will make our research understandable even to those who are not specialists through research on computer-played shogi and the computer-played soccer through the RoboCup. I think that at least part of what we do has to be understandable to anyone.

One sharp-tongued colleague once remarked "You say you'll make Mighty Atom, but in actuality all your work is in shogi and soccer. Not very convincing you know." I suppose it seems that way. But in fact, in order to make Mighty Atom one has got to study a whole raft of the complicated things humans do.

From the Study of Intellect to the Study of Feelings

©You Sung Gil

In 2012, I launched a project to have AI write a novel called "The Whimsical A.I. Project: I'm a Writer." My earlier research on shogi and soccer, and the overseas studies focusing on go and chess have been focused on the study of intellect used to solve very difficult challenges for AI. As we have seen from the recent news of computers beating titled professionals, so the goal of developing AI capable of besting the best human players has been more or less reached. So now it seems to me the time is ripe to launch substantive research on human feelings.

People around me said, "It's too early," "You'll never do it." But I'm the kind of person who isn't really interested in something that can be done if you just try. Also, I and my teammates are convinced that anything a human can do a computer can also do. I've taken a lot of flack for all kinds of research I do, but in fact computers do now play shogi and even soccer.

Our approach to this study begins with the hypothesis that human feelings, creativity, and originality all come from what we call "random generation." You come up with all sorts of options and then you choose the best from among them. When you find something good, it will be praised by others. And we came up with the idea that a computer, too, could write an original story using random generation.

It would be really difficult to write a full-length novel, so we decided to try the short-short type of short story. And we made it our goal to get some kind of prize by submitting the story for the Nikkei Shimbun newspaper-sponsored literary prize Hoshi Shinichi Award.

What a Computer Needs to Write

We soon learned that for a computer to write a novel, it has to be capable of three things: generating natural sentences (in this case Japanese), generating a story, and evaluating whether what is written is interesting.

At present, we have enabled the computer to generate natural sentences. Teammate Sato Satoshi worked out a way to input the basic story using a fixed grammar and a system with which the computer can "write" it into a narrative text. Using that system, we created the first short-short story written by computer.

The other two capabilities--generating the story and evaluating it--are still very difficult. Human beings have to provide the main components of the story and its style, with instructions like "Begin by describing the weather," "Next have the protagonist speak," and so forth. Moreover, the choice of which story to submit to the contest from among the random stories written by the computer, as well as the evaluation of whether or not a story is interesting and how it measures up as a literary work are all done by humans. The computer itself has no idea about what makes interesting content.

In the 2016 Third Hoshi Shinichi Award, one work thus written using AI passed the first screening, and that was given much attention in the media both in and outside Japan.

The "Mind-Reading" Robot

©You Sung Gil

I believe that if robots are to become part of daily life in workplaces and homes, they will have to be made capable of detecting human feelings and emotions. Today, if you say to a robot, "Oh, it's so hot today." It will answer with cold information: "Yes, it is X degrees today. The humidity is X percent." There's nothing very companionable about that. Rather, if it is daytime, the robot might turn on the air conditioner, or if it is evening, it might offer you a beer. If it could read the context and the tone of a human's voice, it could ideally be a robot that could "read your mind"--in the same way a long-time partner will anticipate your needs and desires without you actually saying something.

Understanding Humanity through AI

©You Sung Gil

The history of AI research is a history of failures. You might think you want the computer to think and act like a human being, but it rarely does. AI research--writing programs, correcting them, and so forth--is unostentatious and very time-consuming. Like the "whack-a-mole" game, no sooner do you think you've dealt with one problem than another rears its head.

One thing about this research though: It reminds you how amazing the human being is. I myself think I must be studying AI as a means of understanding humans. I still have a very long way to go, but sometimes I think I have come a little closer. And that is part of the rewards of this research.

Mighty Atom, as you may know, registered worry and puzzlement, realizing that what he thought was beautiful as a robot seemed to be different from what human beings thought was beautiful. That aspect of Atom always impressed me, and I have wanted to make an AI robot capable of self-inspired worry or puzzlement. But that kind of achievement is still far away. We researchers have to comfort and inspire ourselves with what small progress we have made in that direction.

【Interview：February 2017】
Compilation：Yamagishi Hayase