Predictive pursuit emerges in high-dimensional recurrent neural networks.

Area of Science:

• Computational neuroscience
• Animal behavior
• Machine learning

Background:

• Tracking moving objects is vital for survival.
• Neural mechanisms for predictive pursuit are not fully understood.
• Previous studies used rodent models for visual pursuit.

Purpose of the Study:

• To develop a recurrent neural network (RNN) model for predictive pursuit.
• To investigate the computational principles underlying anticipatory behaviors.
• To explore the role of neural network dimensionality in pursuit.

Main Methods:

• Developed a recurrent neural network (RNN) model.
• Trained the model on stereotyped target trajectories.
• Analyzed emergent behaviors and neural representations.
• Ablated specific neural units to test causal roles.
• Trained RNNs with varying dimensionality (rank).

Main Results:

• The RNN model generated internal predictions of target locations.
• Anticipatory behaviors increased with target trajectory exposure.
• Model's pursuit strategy aligned with rodent behavior.
• Units encoding egocentric target position were identified and found crucial for pursuit.
• High-dimensional networks were necessary for anticipatory behaviors, not just pursuit performance.

Conclusions:

• Predictive pursuit emerges in high-dimensional RNNs.
• Egocentric target representation is key for efficient pursuit.
• Network dimensionality is a critical resource for complex predictive behaviors.