Structure of attractors in randomly connected networks

Area of Science:

• Computational neuroscience
• Statistical physics
• Complex systems

Background:

• Randomly connected neural networks exhibit complex dynamics.
• Understanding attractor structures is crucial for neural computation.
• Discrete-time synchronous update rules govern network state evolution.

Purpose of the Study:

• To theoretically analyze the dynamics of randomly connected neural networks.
• To model the evolution of system states over time.
• To characterize the structure of attractors within these networks.

Main Methods:

• Theoretical analysis of deterministic dynamics in binary neural networks.
• Modeling state overlap as a Markovian stochastic process for large networks.
• Utilizing state concentration probability for analytical derivations.

Main Results:

• Demonstrated that state overlap follows a Markovian stochastic process in large networks.
• Developed a model predicting state revisitation frequency based on system size.
• Analytically derived network characteristics, including attractor number and cycle lengths.

Conclusions:

• The Markovian process accurately predicts network behavior and attractor properties.
• Analytical predictions align well with numerical simulations for large networks.
• Provides a theoretical framework for understanding attractor structures in random neural networks.