"The complex workings of the brain, which cannot be seen from the outside, can be clearly deciphered by reflecting them in the 'mirror' of AI. This approach is the greatest fascination and strength of combining brain science and AI."

Professor Makino explains the core of his research in this way. What specifically can be understood by overlapping the biological brain and AI?

For example, in experiments with mouse brains (*1), neural activity is visualized using methods such as calcium imaging (*2). When a mouse performs a specific action, the activity of individual neurons in the brain can be observed through a microscope as they light up.

However, the inside of the brain is extremely complex. Even if we can track the correspondence between "externally visible behavior," such as a mouse moving its limbs, and neural activity, it is difficult to decipher the underlying "perception," "cognition," or the "invisible internal states" such as what and how calculations are being performed in the mind.

This is where "AI agents" created in virtual space come in.

These agents have artificial neural networks (*3) that mimic the neural circuits of the brain, with artificial neurons as their smallest constituent units. Since their activity is pre-designed and defined as a program, it is possible to completely understand and visualize "which unit is performing what calculation."

Professor Makino had an AI agent learn in a virtual space to behave in the same way as an actual biological mouse, and then compared the movements of the artificial neurons with the neural activity patterns of the mouse.

As a result, a phenomenon occurred where the activity of specific artificial neurons and the waveforms and fluctuations of biological neuron activity measured from the mouse brain overlapped with surprising precision.

"For example, we cannot tell what the brain is doing while a mouse is learning just by looking at its behavior. However, by comparing it with the functions of artificial neural network units that represent states and the value of actions in advance, we can gain a dramatically better understanding of what kind of information processing the mouse brain is doing and what it is doing."

In fact, there have been past instances where AI and brain mechanisms have coincidentally matched in this way.

Researchers overseas have already discovered that "dopamine," a neurotransmitter in the brain, and a calculation algorithm called "TD error" (Temporal Difference error) in AI reinforcement learning play exactly the same role.

TD error refers to the "gap (error)" that occurs between the "previously predicted result" and the "actually occurring reality" when an AI learns something.

For example, if you suddenly receive 1 million yen without any warning, a large reward is obtained despite the prior prediction being zero, resulting in a huge "positive error" in the AI system.

In the human brain, this is a pleasant surprise, and at this time, a large amount of dopamine is released. It is a neurotransmitter responsible for motivation, concentration, pleasure, and the coordination of movement.

Conversely, if you expected and predicted that "something good will happen today" but received nothing, you would be disappointed and dopamine release would be suppressed. At this time, a "negative error" occurs in the AI.

What can be inferred from this phenomenon that is similar in both AI and brain science? Professor Makino explains.

"It can be said that a mathematical system (TD error) independently designed by computer scientists on a computer was similarly implemented within the brains of humans and animals. In other words, an algorithm created by humans and the brain system built by nature happened to coincide."

And this phenomenon resonates deeply with the research Professor Makino is currently working on.

"I am very interested to see if AI becomes more 'brain-like' by putting findings obtained from brain research into AI. It is often said that 'AI agents do not do things that make them happy,' but I think it is also important to give AI things that humans have but AI does not."

Identifying the differences that lie between biological brains and AI, and the process of narrowing that gap, leads to the evolution of AI. At the same time, using evolved AI models to understand the brain even more deeply—Professor Makino is trying to create such a "virtuous cycle of knowledge."

How was Professor Makino's own brain, which generates flexible ideas, nurtured?

He grew up in a place rich in nature and was an "ordinary kid" who liked soccer and games. One day, he happened to read a book about the brain and became interested in how the left and right brains have different roles and functions. To study brain science and psychology at university, he went to the University of St Andrews in the UK. Although he couldn't speak much English, he dove into a university overseas.

"I might have had a vague longing for going abroad, but I think I was just wandering even back then. You could say I was curious, or you could say I wasn't settled."

He decided to become a researcher in his third year of university. The catalyst was a one-year exchange program at the University of California, San Diego, where he experienced full-scale research for the first time at the Salk Institute. After graduation, he entered a doctoral program at Cold Spring Harbor Laboratory in the US, where he immersed himself in research on brain synapses and learned the research methods that form his current foundation.

A major turning point was a science magazine he saw in 2015 during his time as a postdoc at the University of California, San Diego. A paper on AI combining reinforcement learning and deep learning by the global AI development company "DeepMind" graced the cover.

The concept of reinforcement learning has a history of originating from findings in psychology and brain science and being applied to computer science, so it is not without connection to the brain science that Professor Makino specializes in. He did not feel that the hurdle of deciphering the mathematical formulas was particularly high.

There were also plenty of educational materials and reproduction codes available online. By teaching himself programming and writing code to run reproduction experiments, he felt a sense of accomplishment, thinking, "This is much more usable than I thought." Artificial neurons in AI and neurons in biological brains—if he compared the activities of these two, he might find a one-to-one correspondence.

"The moment I realized that, I felt that I might be the only one in the world right now who had noticed this possibility and was about to tackle it head-on. I had a sensation like a shiver of excitement."

One of the research themes Professor Makino is currently passionate about is the construction of a "virtual society" using AI and the elucidation of the mechanisms of mental disorders such as depression and stress.

The environment of the Keio University School of Medicine has the great strength of allowing for daily discussions with practicing psychiatrists. Psychiatrists interact with and treat patients through "words," and even though their brains are not directly connected, the invisible exchange of conversation changes the circuits of the other person's brain into a favorable state.

On the other hand, in a state of depression, the characteristic of "no longer seeking rewards that should normally be pleasing" appears. Professor Makino is thinking of reproducing this state using the AI mechanism for calculating the "discrepancy between expectation and result." By comparing it with the "TD error" system and adjusting the AI's calculation settings, he says there is a possibility of creating a "depressed AI" on a computer.

Furthermore, he will conduct experiments in virtual space to see what kind of words from an AI mimicking a psychiatrist's dialogue will help this "depressed AI" recover. If this is realized, it may be possible to elucidate the fundamental mechanism of "why people's minds recover through verbal dialogue" from the perspective of brain science.

"Many people in the world live with intense stress." Feeling this deeply in the city and on crowded trains, Professor Makino is searching for the ideal form of a "kind society" where everyone can be fulfilled both mentally and physically.

On the other hand, it is not easy to eliminate excessive stress. The source of stress is the "gap between expectation and reality," and while giving up on ideals reduces stress, it may also lose the "competition" that generates social development and vitality.

What is expected there is a "digital virtual society" where thousands or tens of thousands of AI agents live in virtual reality. Through digital society simulations originating from brain science, he searches for hints for an "optimal balance" that can achieve both moderate competition and low stress.

When training an AI, it is known that rather than intensive training in only one thing from the beginning, having it accumulate many seemingly unrelated "various experiences and trials and errors" results in an overwhelmingly higher growth speed and potential when it is finally narrowed down to a specialized field.

From the perspective of brain science as well, accumulating a variety of behavioral and experiential patterns while young significantly increases subsequent adaptability and learning efficiency. Rather than limiting one's possibilities from the start, taking a detour on purpose leads to a great blossoming of talent in the long run.

Professor Makino himself has not followed a straight path but has blossomed in his unique research field as a result of repeated detours and explorations.

"I hope that you young people will not fit yourselves too much into a future mold, but will first freely expand your interests toward the 'things you want to know' that are close to you. The accumulation of those experiences should surely bring about an unexpectedly large leap in the future."

Composition: Toru Tamakawa, Editor-in-Chief of Asahi Shimbun GLOBE+

Interview and text: Kaori Sawaki

Photography: Hiromi Shinada