A-SOID

A-SOID performs active, expert-guided, data-efficient discovery and classification of naturalistic behaviors by combining supervised and unsupervised learning for behavioral segmentation and identification across species.

Key Features:

• Hybrid Learning Approach: Combines supervised and unsupervised learning to iteratively learn user-defined groups and expand classification with reduced labeled data.
• Data Efficiency: Demonstrated an 85% reduction in required training data in experiments with socially-interacting mice while maintaining superior performance versus other methods.
• Unsupervised Classification Capabilities: Identifies and isolates ethologically distinct interactions not initially apparent, including the discovery of two additional distinct mouse interactions in the reported study.
• Transparency and Interpretability: Uses a game-theoretic approach to reveal defining features of supervised classifications and clarify cluster definitions.
• Versatility Across Species and Modalities: Demonstrated performance on non-human primates using 3D pose data and applicable to human behavior segmentation and single-person behaviors using 3D position keypoints.

Scientific Applications:

• Animal behavior studies: Analysis and segmentation of complex social interactions and behaviors in model organisms such as mice.
• Ethology and behavioral segmentation: Discovery and classification of ethologically distinct subactions and interactions.
• Human social interaction analysis: Segmentation and classification of social and single-person human behaviors using 3D position keypoints.
• Cross-species behavioral comparison: Comparative analysis of behaviors across mice, non-human primates, and humans using 3D pose/keypoint data.

Methodology:

Active learning framework integrating supervised and unsupervised learning, iterative learning of user-defined groups, unsupervised classification/clustering to discover novel interactions, a game-theoretic approach to reveal defining features of supervised classifications, and use of 3D pose/3D position keypoint data.