Dual-Stream Spatiotemporal Networks with Feature Sharing for Monitoring Animals in the Home Cage.

Area of Science:

• Computational Neuroscience
• Machine Learning
• Animal Behavior Analysis

Background:

• Accurate classification of animal behavior is crucial for understanding neurological and physiological states.
• Traditional methods often struggle with the complexity and subtlety of spatiotemporal mouse movements in home-cage environments.

Purpose of the Study:

• To develop and evaluate a novel spatiotemporal deep learning approach for enhanced mouse behavioral classification.
• To introduce and validate a 'feature sharing' technique within dual-stream neural network architectures.

Main Methods:

• Utilized a series of dual-stream deep learning architectures with modifications for optimal performance.
• Implemented a novel 'feature sharing' approach for joint stream processing at regular intervals.
• Ensembled Inception-based and attention-based networks incorporating feature sharing for improved classification accuracy.

Main Results:

• Feature sharing architectures consistently outperformed conventional dual-stream networks with standalone streams across all models.
• Inception-based architectures demonstrated significant accuracy gains (6.59%–15.19%) due to feature sharing.
• Ensembled models achieved superior classification accuracy, validated on additional mouse behavioral datasets.

Conclusions:

• The proposed feature sharing deep learning approach offers a significant advancement in mouse behavioral classification.
• This method provides a robust and accurate tool for analyzing complex animal movements in ecological settings.
• The developed models show promise for broader applications in behavioral neuroscience research.