Combining AI and Neuroscience to Enhance Understanding of Animal Behavior

In recent years, artificial intelligence (AI) has become a hot topic across academia and industry. Its expanding applications now include analyzing the complex movements of animals.

The Center for Cognition and Sociality within the Institute for Basic Science (IBS) has developed an AI-based analysis tool called SUBTLE, which can classify and analyze animal behavior by learning from 3D motion data. This collaboration between neuroscience and data science has enabled AI to accurately distinguish complex animal movements. The integration of AI into animal behavior analysis, a field with potential implications for human disease research, has significantly improved the precision of such studies.

Dr. KIM Seonpil, a postdoctoral researcher who participated in the study, remarked, “This research has given me confidence that integrating AI into neuroscience can achieve what was previously impossible.” Fellow postdoctoral researcher Dr. KWON Jae added, “Ultimately, creating a big data repository of behavioral data from various animals could help us uncover fundamental principles governing behavior across all biological species.”

Here, the two researchers share insights into their past achievements and future aspirations.

Q. Please introduce yourself.
KIM Seonpil
I am KIM Seonpil, a postdoctoral researcher at the Center for Cognition and Sociality. I completed my bachelor's degree in biology at Yonsei University and joined Dr. C Justin LEE’s lab as a student intern in 2014 before pursuing an integrated master’s and Ph.D. program. During my studies, I researched dopamine signaling pathways in astrocytes and their role in social memory. Since 2023, I have been conducting various research projects at IBS as a postdoctoral researcher.
KWON Jae
I am KWON Jae, a researcher focused on AI-based neuroscience. I earned my Ph.D. in 2022 from Korea University under the guidance of Dr. C Justin LEE at the IBS Center for Cognition and Sociality. I worked as a postdoctoral researcher at IBS until September 2024. Starting in November, I began a new postdoctoral position with Dr. Cha Meeyoung at the Max Planck Institute, whom I met through collaborative research during my time at IBS.

Q. Please explain your main research responsibilities at IBS.
KIM Seonpil
I have diverse research interests, including astrocyte functions, visual information-based social memory, and the discovery of low-toxicity dopamine variants. The development of SUBTLE emerged from my desire to analyze mouse behavioral experiments more objectively and efficiently.
KWON Jae
My work focuses on the intersection of machine learning and neuroscience. This involves either applying machine learning techniques to neuroscience research or advancing machine learning inspired by neuroscience.

Q. Please introduce some of the most notable achievements of the Center for Cognition and Sociality.
KIM Seonpil
I’d highlight our SUBTLE paper as one of the center’s key achievements. Since AI research draws heavily on the structural organization of neural cells in the brain, neuroscience and AI are closely connected. Although AI-based research in neuroscience is growing, our SUBTLE study is the first example from our center to employ AI.
KWON Jae
I agree that the SUBTLE algorithm, developed earlier this year, is a landmark achievement. By applying our machine learning algorithm, we’ve streamlined one of the most labor-intensive tasks in neuroscience: animal behavior analysis. This advancement is expected to deepen our understanding of animal behavior.

Q. How did the development of SUBTLE begin?
KIM Seonpil
The idea for SUBTLE was sparked by our introduction to a device called AVATAR, which reconstructs mouse movements in 3D. While the device could generate 3D motion data, we lacked a tool to analyze it. To address this, we adopted unsupervised learning to analyze the data objectively. We reduced the dimensions of the high-dimensional data and clustered it, identifying distinct behaviors within each cluster. Ultimately, SUBTLE demonstrated its ability to analyze a range of movement datasets over time.
KWON Jae
We started this project to develop a more efficient and objective way to analyze animal behavior. Although AVATAR could collect vast amounts of data, distinguishing meaningful differences in behavior—such as walking, standing, grooming, or pausing—was a challenge. We used unsupervised machine learning to provide a practical solution for this issue.

Q. SUBTLE has been applied to analyze mouse behavior and shown potential for use with humans and primates. What other animals could it be applied to?
KIM Seonpil
One of the key principles behind SUBTLE is its universality. Theoretically, it could analyze the movements of any animal—and even non-living objects. We expect it to be applicable across a wide range of scenarios.
KWON Jae
SUBTLE is not limited to any specific species. As long as the subject has “key points” that can be tracked, such as fish or humanoid robots, their movements can be analyzed. Additionally, SUBTLE has the potential to analyze not just individual subjects but also group behavior patterns.

Q. How is SUBTLE being received in academia? We would like to know potential applications of animal behavior categorization and analysis.
KIM Seonpil
One of SUBTLE’s most significant contributions is its introduction of a benchmarkable metric called the Temporal Proximity Index (TPI). Benchmarks are crucial in AI research for developing and testing new algorithms. I anticipate that our work will lead to the creation of more precise and faster behavior analysis tools.
KWON Jae
If successfully applied, SUBTLE could make significant contributions to fields like veterinary medicine, animal welfare, ecology, conservation research, robotics, and AI. Moreover, big data on animal behavior could help us uncover fundamental principles governing behavior across species. This kidn of research can redefine the framework of animal behavior analysis and deepening our understanding of life.

Q. Minimizing human intervention seems crucial in animal behavior analysis. What challenges did you face during this research?
KIM Seonpil
Traditional behavior analysis often relies on human involvement, which can introduce variability and requires significant time and effort. Even AI-based analysis often involves manual adjustments to “hyperparameters,” which is another form of human intervention. We worked hard to minimize human input in SUBTLE to ensure objectivity in our analyses.
KWON Jae
The biggest challenge was enabling the algorithm to classify behaviors in a way that aligns with human expectations without direct human input. Essentially, we had to resolve the question: “What constraints should we impose so the classification resembles human logic?” Answering this was the toughest part of the research.

Q. What kind of support would help your future research plans?
KIM Seonpil
This study has strengthened my belief that integrating AI into neuroscience can achieve unprecedented outcomes. I’m particularly interested in exploring the application of AI in social behavior research, a relatively untapped field.
KWON Jae
Currently, SUBTLE focuses on individual behavior analysis, but I aim to extend its application to multi-animal scenarios. By doing so, we could analyze and interpret complex interactions between individuals, which would significantly contribute to social behavior research.