Internal dynamics of recurrent neural networks trained to generate complex spatiotemporal patterns

Area of Science:

• Computational Neuroscience
• Machine Learning
• Complex Systems

Background:

• Understanding neural system pattern generation is crucial for both neuroscience and machine learning.
• Investigating the link between individual neuronal activity and complex network behaviors is a key challenge.

Purpose of the Study:

• To elucidate the microscopic features underlying spatiotemporal pattern generation in recurrent neural networks.
• To analyze the role of individual neural trajectories and network activity distributions in creating specific dynamic regimes.
• To examine the contribution of trained output weights to autonomous multidimensional dynamics.

Main Methods:

• Utilizing recurrent neural networks with feedback, specifically reservoir computers.
• Analyzing microscopic features of neuronal activity.
• Investigating spatiotemporal pattern generation, including multicluster and chimera states.
• Examining individual neural trajectories and whole-network activity distributions.

Main Results:

• Identified microscopic features responsible for generating complex spatiotemporal patterns.
• Demonstrated the influence of individual neural trajectories on network dynamics.
• Showed how network activity distributions contribute to specific dynamic regimes like multicluster and chimera states.
• Investigated the impact of trained output weights on multidimensional dynamics.

Conclusions:

• Recurrent neural networks, particularly feedback reservoir computers, can generate complex spatiotemporal patterns.
• Individual neuronal activity and network-level dynamics are critical determinants of emergent network states.
• Understanding these mechanisms offers insights into neural computation and advanced machine learning models.