Reservoir computing in simulated neuronal cultures: Effect of network structure.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Computational Neuroscience

Background:

• Biological neurons offer energy efficiency and self-repair for AI.
• Real neurons exhibit variable, noisy dynamics and spontaneous activity.
• Neuronal networks have complex, plastic, and evolving connectivity.

Purpose of the Study:

• Investigate input response stability in neuronal cultures using reservoir computing.
• Analyze the impact of network topology (homogeneous vs. modular) on performance.
• Assess effects on fading memory, noisy dynamics, and rewiring robustness.

Main Methods:

• Utilized a numerical model of neuronal culture growth and activity.
• Employed a reservoir computing framework.
• Compared homogeneous and modular network architectures.

Main Results:

• Modular networks showed longer fading memory.
• Modular networks sustained higher performance under noisy conditions.
• Modular networks were more robust to connectivity rewiring.
• No correlation found between spectral properties of adjacency matrix and reservoir performance.

Conclusions:

• Modular network topology enhances stability and performance in neuronal cultures for AI.
• Network architecture is critical for robust reservoir computing with biological neurons.
• Spectral properties do not predict reservoir computing capabilities in these models.