Students who are selected by the Japan Society for the Promotion of Science (JSPS) for a Research Fellowship for Young Scientists receive research subsidies and/or fellowship grants (Grants-in-Aid) for a period of two to three years. The research grants can be used for any purpose and no reporting is required. On the other hand, fellows can receive a maximum of 1.5 million yen per year by applying for a Grant-in-Aid for Scientific Research. Of course, JSPS Research Fellowships for Young Scientists are highly competitive, so not everyone who applies will be accepted. For this article, we sat down with Hayata Kaneko, a first-year doctoral student in the Graduate School of Science and Engineering who advanced to the graduate school on the Early Enrollment Program and was then selected for JSPS DC1 Research Fellowship.

Skipping ahead to M1 from UG3

—— How did you end up advancing to the master’s program on the Graduate School Early Enrollment Program?

Kaneko: When I was a third-year undergraduate student, I was approached by Professor Meng, whose laboratory I was studying in. I was surprised because I didn't even know about the early enrollment program at the time (laughs). Naturally, I was happy to be asked, but I was also worried about what would happen afterwards, so I was not sure if I wanted to apply at first. Because, after all, I would be suddenly thrust into the first year of a master’s program (M1). Moreover, if you plan to get a job after graduating with a master's degree, you have to start thinking about your career path from about partway through M1. However, after talking with many different faculty members I came to realize that it was a great opportunity and decided to give it my best shot.

—— After that, you decided to advance to the doctoral program instead of getting a job with a master's degree.

Kaneko: If you want to find a job, you need to start doing internships and other activities as soon as you enter M1. At the time, I was worried that if I did that, I might not be able to produce any kind of satisfactory results because I would not have enough time to do research. So, thinking I had to produce some kind of research results as soon as possible, I devoted myself entirely to my research because I had been accepted on the early enrollment program. Then, before I realized it, I was captivated by how fun research is, so I decided to advance to the doctoral program to deepen my research.

—— In 2022, you had already published a paper with Dr. Meng by October of M1.

Kaneko: I am interested in AI and wanted to try my hand at using software to create images. Therefore, we tried, probably for the first time in Japan, to automatically repair degraded characters using a generative adversarial network(GAN). I co-authored a paper summarizing the results of the project, "An Attempt to Restore Japanese Classical Texts Using a Generative Adversarial Network," with Dr. Meng, which was published in ART RESEARCH (Art Research Center of Ritsumeikan University), Vol. 23-1. AI-driven restoration of ancient documents and the transmission of cultural heritage is also the topic I chose when I later applied for the JSPS fellowship.

Transforming from a game enthusiast into a researcher

—— Had you been interested in ancient documents for some time?

Kaneko: It was not something that I would say I particularly liked (laughs). As a child, I was interested in pyramids and such, as well as ancient characters. But to be honest, I was always a gamer, and I was hooked to the point of making games myself. I even thought that if I made my own game and took it to a game developer, I might be able to get a job with them (laughs). However, it takes a great deal of time to carefully create the characters and the scenario, so I have only been able to make one complete game. Looking back on it now, I realize that making games was similar to research in that it was a process of deciding your own goals and working diligently to achieve them.

—— You co-authored a paper with your older labmates in 2023.

Kaneko: Yes, that would be “Deteriorated Characters Restoration for Early Japanese Books Using Enhanced CycleGAN”, Heritage 2023, 6, 4345–4361. (https://doi.org/10.3390/heritage6050230). We wrote about using a GAN to restore ancient Japanese documents. Using image generation to restore documents is rewarding because it you get tangible results. When we were able to clearly restore the previously illegible characters, we were filled with joy. Going forward, more data will be needed to improve the accuracy of restorations, but there is a limited amount of data to begin with when it comes to ancient texts. Devising ways to achieve this is one of the foundations of my current research.

—— Do you have an approach in mind for how you can solve this problem?

Kaneko: I have come to think that this problem is rather difficult to solve with software alone. Dr. Meng approaches his research through the dual lenses of software and hardware, but I initially told him that I wanted to pursue my research with a software-centered approach. However, I came to realize that I really needed to do something in terms of hardware in order to continue my research. This is why I spent the entire second year of my master's program thoroughly learning the fundamentals of hardware.

Striving to make AI more powerful and faster

—— Did you find it difficult to switch gears from researching software to learning about hardware?

Kaneko: I learned by reading reference books, so in other words, I taught myself. Even so, nowadays there is all kinds of information, including open-source information, available on the internet that you can refer to while assembling your own hardware. Of course, it would have been a tremendous challenge if I had tried to build it all by myself from scratch.

—— So, you even built your own hardware for the purpose of your research?

Kaneko: That's right. How compactly and quickly you can process data makes all the difference between success and failure in AI development. Among AIs, neural network calculations use the multiply-accumulate operation. This operation computes the sum of the products of corresponding elements for two vectors. However, since this requires an enormous amount of computing power, parallel operations, which process multiple calculations simultaneously, are utilized to speed up the process. The processors used in this process are GPUs, the graphic processing units made famous by NVIDIA, but they are very expensive and not readily available. For this reason, I am trying to devise some way to create hardware that can compute as fast as possible on a small scale, using only CPUs.

—— But isn’t building hardware yourself a lot of work?

Kaneko: To be honest, it is a very time-consuming process. After all, one bug can cause the computer to stop, and when that happens, your only option is to start debugging. So, recently, I have been spending my research time on this debugging process. Still, the satisfaction of coming up with new AI architecture on my own, adding new computing functions to the hardware, and getting the expected results is a powerful motivator to keep moving ahead. I am now running repeated simulations to check the hardware, and it is getting much faster. I have a conference in less than two months (scheduled for the end of March 2025), so I am working hard to finish it by then.

Earning the JSPS Fellowship for large-scale research that combines software and hardware

—— What topics are you working on now?

Kaneko: I have been able to enhance the computing functions of my hardware, but that in and of itself is not the original goal. Having searched high and low for something only I can do, I am currently working on developing a method to speed up the process by approaching the issue from the software side as well. A huge number of papers are published about software every day, so I am incorporating this knowledge into my research by referring to sites that summarize papers published in top journals. This doesn't mean I spend the entire day doing nothing but research, however.

—— How do you typically spend your time?

Kaneko: I’d say I do research for about eight hours. There is a limit to my concentration, so when I get home, I relax by watching movies and videos. However, once I step foot in the lab, I make a conscious effort to sustain my concentration. Also, I try to never give up in the planning stage, and I am conscious of the fact that I should not overthink things and just get started. Even though this is a field that I think I like, it is also a domain where you can’t know anything until you try it.

—— What points did you pay attention to when writing the application form for the JSPS Fellowship?

Kaneko: What I paid the most attention to was the scale of the research. Since this is a three-year research project, I was conscious about conveying a reasonable sense of scale. Of course, since I would be devoting three years of my mid-twenties to this project, if I may exaggerate, I also felt that I was making a decision that would influence the trajectory of my life. I think I came to think that way because of the many international students around me. Looking at them, you can see that they are truly putting their lives on the line. Dr. Meng also corrected my application many times. Another good thing was that I was able to submit a paper for publication during my master's program. I was able to use the figures and other materials that I used in that paper in my application, so I think that allowed me to communicate my research more clearly.

—— What do you have in mind for your future career path?

Kaneko: While I would like to work on both software and hardware like Dr. Meng, right now, I just want to devote myself to hardware architecture. After I earn my degree, I would like to work in research, preferably not in a company but in an academic institution like a university.

5