Identifying behavioral structure from deep variational embeddings of animal motion

Area of Science:

• Neuroscience
• Computational Biology
• Machine Learning

Background:

• Analyzing hierarchical behavioral organization is crucial in neuroscience.
• Markerless pose estimation visualizes complex animal motion dynamics.
• Robust methods are needed to segment and structure behavioral data.

Purpose of the Study:

• To develop an unsupervised deep learning framework for behavioral structure identification.
• To segment animal motion into discrete, hierarchically organized motifs.
• To analyze behavioral dynamics in a mouse model of beta amyloidosis.

Main Methods:

• Developed a deep variational embedding framework for animal motion (VAME).
• Employed unsupervised probabilistic deep learning for behavioral analysis.
• Applied VAME to a mouse model of beta amyloidosis.

Main Results:

• VAME successfully identified discrete behavioral motifs and their hierarchical usage.
• The framework grouped motifs into communities, revealing cohort-specific usage differences.
• Detected subtle behavioral changes in mice undetectable by human observation.

Conclusions:

• VAME offers a robust approach for segmenting animal motion and uncovering behavioral structure.
• The framework enables hierarchical analysis of behavioral motifs across diverse experimental conditions.
• This method advances the objective quantification of complex behaviors in neuroscience research.