A-SOiD, an active-learning platform for expert-guided, data-efficient discovery of behavior.

Area of Science:

• Ethology
• Machine Learning
• Behavioral Science

Background:

• Supervised and unsupervised methods are popular for naturalistic behavior identification.
• Each method has limitations, including user bias, training cost, complexity, and action discovery challenges.

Purpose of the Study:

• To introduce A-SOiD, an active-learning platform that integrates supervised and unsupervised learning.
• To overcome the inherent drawbacks of existing behavior identification methods.
• To provide an efficient and intuitive tool for behavior segmentation and discovery.

Main Methods:

• A-SOiD employs an active-learning approach to iteratively learn user-defined groups with reduced training data.
• It utilizes directed unsupervised classification for expansive behavioral categorization.
• A game-theoretic approach enhances transparency in cluster definitions for supervised classification.

Main Results:

• A-SOiD outperformed standard methods in identifying social behaviors in mice, requiring 85% less training data.
• The platform successfully isolated ethologically distinct mouse interactions using unsupervised classification.
• Similar performance and efficiency were observed with nonhuman primate and human 3D pose data.

Conclusions:

• A-SOiD effectively blends supervised and unsupervised learning to enhance naturalistic behavior identification.
• The platform demonstrates significant efficiency and performance gains across multiple species.
• A-SOiD offers an open-source, intuitive interface for behavior analysis and discovery.