The Multi-Agent Behavior Dataset: Mouse Dyadic Social Interactions

Area of Science:

• Behavioral Neuroscience
• Computational Biology
• Machine Learning

Background:

• Understanding complex interactions between agents is crucial in behavioral neuroscience.
• Automated analysis of animal behavior is essential for accelerating research.
• Existing datasets may lack the scale or diversity needed for robust model training.

Purpose of the Study:

• Introduce the Caltech Mouse Social Interactions (CalMS21) Dataset for multi-agent behavior modeling.
• Provide benchmarks for evaluating automated behavior classification methods.
• Facilitate research on inter-annotator differences and learning new behaviors.

Main Methods:

• Collected trajectory data from freely behaving mice in a resident-intruder assay.
• Recorded 6 million frames of unlabeled tracked poses and over 1 million frames with annotations.
• Developed benchmarks for training, style transfer, and few-shot learning of behaviors.

Main Results:

• The CalMS21 dataset offers a large-scale resource for multi-agent behavior analysis.
• Benchmarks address challenges in classifying behaviors from labeled and unlabeled data.
• The dataset facilitates generalization to new behavioral settings and annotator styles.

Conclusions:

• The CalMS21 dataset is a valuable tool for advancing automated behavior classification in neuroscience.
• It supports the development of more robust and generalizable multi-agent behavior models.
• This resource will accelerate discovery in social behavior research.