Deep learning improves automated rodent behavior recognition within a specific experimental setup.

Area of Science:

• Neuroscience
• Pharmacology
• Artificial Intelligence
• Computer Vision

Background:

• Automated rodent behavior analysis is crucial for neuroscience and pharmacology research.
• Current automated systems lack adaptivity and can be improved with AI.
• Deep learning offers potential for advanced behavioral recognition.

Purpose of the Study:

• To compare a conventional rat behavior recognition (RBR) system with a deep learning method.
• To evaluate the performance of a multi-fiber network (MF-Net) for behavior recognition.
• To assess the transferability of performance across different experimental setups.

Main Methods:

• Implementation of a deep learning model, MF-Net.
• Application of data augmentation strategies, including video cutout and dynamic illumination change.
• Comparative analysis of MF-Net against a conventional RBR system within and across experimental setups.

Main Results:

• MF-Net, with data augmentation, significantly improved within-setup performance compared to the conventional RBR system.
• Performance gains achieved within a specific experimental setup did not transfer to other setups.
• Novel video augmentation techniques (video cutout, dynamic illumination) were introduced.

Conclusions:

• Deep learning models like MF-Net show promise for enhancing automated rodent behavior analysis.
• The lack of cross-setup performance transfer highlights challenges in generalizing AI models for behavioral research.
• Further research is needed to address model adaptability and improve generalization across diverse experimental conditions.