Shenzhen News Network, January 10, 2024 (Reporter Ye Mei) What does it mean for animals to exhibit complex social behaviors in their daily lives? How do birds silently convey information when they flock and take off synchronously? Why do fish change their swimming patterns when threatened? Exploring the neural mechanisms and behavioral patterns behind these questions can help understand the deeper meaning of animal social language.
The development of artificial intelligence technology has not only greatly helped scientists analyze and understand biological behavior, but also provided a revolutionary perspective for neuroscience research.
At 6 p.m. on January 8, Beijing time, the latest research results of the research team of the Institute of Brain Cognition and Brain Diseases of the Shenzhen Institute of Advanced Technology of the Chinese Academy of Sciences and the researcher team of Wei Pengfei of Shenzhen University of Technology (in preparation) were published in the top international academic journal "Nature - Machine Intelligence". The research team applied artificial intelligence technology to neuroscience research and proposed a small-shot learning Xi computational framework model for studying social behavior, the Social Behavior Atlas (SBeA), which realized the label-free, high-precision three-dimensional pose estimation, zero-shot identity recognition and refined social behavior classification of free social animal models, solved many important difficulties in accurately detecting animal social behavior, and is expected to innovate the research paradigm of social behavior neural circuit mechanism.
Team Leader: Researcher Wei Pengfei
Prof. Wei Pengfei is the corresponding author of the paper, and Han Yaning, a doctoral student, is the first author of the paper. The co-first authors include Chen Ke, a master's degree graduate, and Wang Yunke, a former assistant researcher at the Institute of Brain Research. Researcher Wang Liping, director of the Institute of Brain Research of Shenzhen Advanced Institute, gave important guidance and help in the research process. Shenzhen Advanced Institute is the first unit of the paper.
"Multi-animal behavior quantification is the key to interpreting animal social behavior, and has a wide range of applications in neuroscience and ecology. Trenton Jerde, senior editor of the journal Nature Machine Intelligence, spoke highly of the study.
AI-empowered, zero-label, accurate identification of animal identity
In the free social state, animals with different disease models often have different disease states, and related behavioral phenotypes need to be mapped one by one. However, the animal's exterior is completely covered with fur, making it difficult for even experienced professional experimenters to distinguish between the identity and status of each animal in a social setting.
Schematic diagram of the tandem research of artificial intelligence to help neurological diseases from the laboratory, natural environment to clinical care Courtesy of the research team
In recent years, the wide application of artificial intelligence technology in the field of traditional behavioral research has promoted the vigorous development of computational neurobehavior, and AI animal behavior tracking technologies such as DeepLabCut, SLEAP, and MoSeq are becoming important research tools for neuroscientists.
However, when analyzing multiple animal targets and animals' free social behavior, the above-mentioned technologies still cannot achieve the identity accuracy of massive data annotation and continuous tracking. At present, there are still no effective tools available to detect animal behavior data, and most of the research on animal social behavior is still in the stage of traditional three-box behavior experiments.
In this study, the research team applied artificial intelligence technology to animal identity recognition, proposed a two-way transfer learning Xi computing framework, and realized the zero-sample multi-animal social identity recognition of the artificial intelligence model without labeling animal identity data in advance, with a recognition accuracy of more than 90%, which fully meets the accuracy requirements of social experiments.
Xi In the non-social scene, it is very simple to distinguish the identity of each animal, and the animal identity information that these models have recognized can be transferred to the multi-animal social scene, which realizes the knowledge sharing and 'brain supplement' between different models, thus solving the problem that the artificial intelligence model needs to manually annotate a large amount of data to achieve multi-animal identity recognition. Wei Pengfei, the corresponding author of the paper, said.
Cross-species universal adaptation to facilitate neurological disease research
In previous research, Wei Pengfei's team developed a single-animal fine behavior analysis algorithm framework and a miniaturized free-moving eye tracking device, which were published in journals such as Nature Communications and Molecular Psychiatry and widely used by researchers at home and abroad.
Group photo of the team, Wei Pengfei (corresponding author, third from left in the front row), Han Yaning (first author, second from left in the front row), Wang Nan (co-author, fourth from left in the front row), Wang Zhouwei (co-author, fifth from left in the front row) Courtesy of the research team
However, these methods can only be applied in single-animal experimental scenarios, which limits the expansion of research from the laboratory to clinical care. Wei Pengfei introduced. After three years of unremitting efforts, the interdisciplinary team led by Wei Pengfei has solved a large number of artificial intelligence technical problems, and combined with the real experimental demand scenario optimization algorithm framework, completed the fine behavior analysis research of multi-animal targets, which is expected to become an important puzzle in the field of computational neurobehavior.
The SBeA technology developed by the research team realizes the parallel, dynamic, and hierarchical decomposition of animal social behavior, and conducts adaptive unsupervised clustering of social behavior characteristics of mice, rats, birds, dogs, non-human primates and other animals, and obtains more than 100 kinds of fine social behavior modules, including chasing, mutual grooming, and aggression.
This method does not need to define social behavior categories in advance, which is conducive to discovering new and undefined social behavior differences, and can identify free social behavior phenotypes that are difficult to obtain in classical social experiment paradigms such as three-box social and sub-regional socialization.
"In the course of our research, we discovered an interesting phenomenon that mice, like humans, will 'actively' care for other mice, especially those in the model of mental illness. This system found that normal mice may have a higher preference for autism model mice. According to Han Yaning, the first author of the paper, this provides a new idea for the team to further analyze the deeper meaning of animal social behavior.
In addition to mice, this method is also suitable for calculating precise 3D social postures, identities, and fine social modules in birds and domestic dogs, with the potential for cross-species applications. Among them, the research work of domestic dogs is mainly completed with the support of the major project of Science and Technology Innovation 2030 "Brain Science and Brain-like Research" in cooperation with the team of academician Zhang Yaping and researcher Wang Guodong of Kunming Institute of Zoology, Chinese Academy of Sciences, and Li Jing of Kunming Police Dog Base of the Ministry of Public Security.
Luo Minmin, co-director of the Beijing Center for Brain Science and Brain-like Research and a professor at Tsinghua University, commented on the results, saying that the Social Behavior Atlas can be regarded as a "magnifying glass" in the field of behavior research, which can help us observe and understand the complex social interactions, behavior patterns and neural basis of animals in a more nuanced way.
"In recent years, the development of computational neurobehavior has revolutionized the classical behavioral paradigm, enabling behavioral observation from the laboratory environment to the natural environment. In the future, AI-enabled neuroscience research will further deepen the understanding of big data physiological parameters from animal models to clinical treatment, provide guidance for the implementation of more accurate and individualized non-invasive neuromodulation, and is expected to help humans break through the 'cage' of understanding complex mental diseases." Wei Pengfei said.
Link to the paper: https://www.nature.com/articles/s42256_023_00776_5 Source: Shenzhen News Network