Modeling the behavior of multiple subjects using a Cauchy-Schwarz regularized Partitioned Subspace Variational...

Area of Science:

• Neuroscience
• Computational Biology
• Behavioral Science

Background:

• Accurate modeling and quantification of behavior are crucial for understanding brain function.
• A significant challenge lies in unifying behavioral modeling across diverse subjects.
• This requires separating common and subject-specific behavioral features.

Purpose of the Study:

• To develop a novel method for partitioning behavioral data into common and distinct features across subjects.
• To model subject-specific behavioral variations using a new regularization technique.
• To uncover the relationship between neural activity and observed behavior.

Main Methods:

• Utilized a semi-supervised approach based on the Partitioned Subspace Variational AutoEncoder (PS-VAE).
• Introduced a novel regularization method employing Cauchy-Schwarz divergence to model distinct behavioral features.
• Developed the Cauchy-Schwarz regularized Partitioned Subspace Variational AutoEncoder (CS-PS-VAE) model.

Main Results:

• The CS-PS-VAE model successfully captures continuously varying behavioral differences among subjects.
• The model effectively identifies distinct behavioral features in an unsupervised manner.
• Demonstrated success in linking recorded neural data with subsequent behavioral patterns.

Conclusions:

• The CS-PS-VAE offers a robust framework for advanced behavioral quantification.
• This approach enhances the understanding of individual variability in behavior.
• The method provides a powerful tool for exploring neural-behavior relationships.