Predicting the long-term collective behaviour of fish pairs with deep learning

Area of Science:

• Computational ethology
• Animal behavior science
• Machine learning applications

Background:

• Collective behavior in animal societies is crucial for survival and reproduction.
• Analytical models are traditionally used to study collective motion.
• Understanding social interactions is key to predicting group dynamics.

Purpose of the Study:

• To introduce and evaluate a deep learning model for assessing social interactions in fish (Hemigrammus rhodostomus).
• To compare the performance of the deep learning model against experimental data and a state-of-the-art analytical model.
• To develop a systematic methodology for validating collective motion models using spatio-temporal observables.

Main Methods:

• Development of a deep learning model to analyze fish social interactions.
• Comparison of deep learning model outputs with experimental observations.
• Benchmarking against a leading analytical model for collective motion.
• Utilizing individual and collective spatio-temporal observables for model validation.

Main Results:

• The deep learning model demonstrates comparable performance to analytical models in reproducing experimental observables.
• Machine learning models can effectively capture subtle, short- and long-term dynamics of collective behavior.
• The deep learning approach shows scalability to larger groups and adaptability to other fish species without retraining.

Conclusions:

• Deep learning offers a powerful, complementary approach to analytical models for studying collective motion in animal groups.
• Consistent validation across timescales is essential for assessing the faithfulness of collective motion models.
• The proposed deep learning framework provides a robust tool for analyzing complex social interactions in animal societies.