The research team of Shenzhen Institute of Advanced Technology and Brain Diseases, Shenzhen University of Technology (SBeA) applied AI technology to neuroscience research, and proposed a small-sample learning computing framework model (SBeA).On January 8, the paper was published in Nature-Machine Intelligence.
In the three stages of attitude estimation, identification and behavior classification of free social animals, SBeA proposes a small sample data generation algorithm, zero-sample bi-directional transfer learning method and unsupervised social behavior decomposition clustering strategy.
According to reports, the design of attitude estimation algorithms for small sample data generation and two-way transfer learning methods for identifying zero samples are inspired by the brain's working mechanisms.The researchers propose a continuous masking copy-paste data algorithm that can effectively synthesize a large amount of new data and train more accurate models to obtain more accurate 3D social posture estimates.SBeA's multi-animal social identification enables two-way transfer learning by sharing knowledge from different networks, eliminating the process of labeling data, and achieving high-precision animal identification with more than 90% accuracy.SBeA does not need to define social behavior categories in advance, facilitates the discovery of new, undefined social behavior differences, and identifies free social behavior phenotypes that are difficult to obtain in classic social experimental paradigms.
In addition to mice, SBeA is also suitable for calculating accurate 3D social postures, identities, and fine social modules for birds and domestic dogs, with potential for cross-species applications."Multi-animal behavior quantification is key to interpreting animal social behavior and has broad application value in neuroscience and ecology.Trenton Jerde, senior editor of the journal Nature-Machine Intelligence, commented.