Excitatory/inhibitory balance emerges as a key factor for RBN performance, overriding attractor dynamics.

Area of Science:

• Computational neuroscience
• Artificial intelligence
• Complex systems

Background:

• Reservoir computing offers efficient computation, often optimized at critical regimes ('edge of chaos').
• Understanding reservoir-to-reservoir variability and network design is crucial for physical reservoir computers.
• Random Boolean Networks (RBNs) are a model system for studying complex dynamics.

Purpose of the Study:

• To investigate the link between connectivity, dynamics, and computational performance in RBNs.
• To determine how network parameter distributions affect dynamics near critical points.
• To evaluate the influence of attractor properties versus network parameters on task performance.

Main Methods:

• Analysis of Random Boolean Networks (RBNs) with varying distribution parameters.
• Identification and statistical quantification of dynamical attractors.
• Evaluation of reservoir performance on memorization and prediction tasks.

Main Results:

• Specific RBN distribution parameters yield diverse dynamics near critical points.
• Most reservoirs exhibit a dominant attractor, but its intrinsic dynamics showed little performance impact.
• A positive excitatory balance optimized memory performance at a critical point.
• A negative inhibitory balance optimized prediction performance at a different critical point.

Conclusions:

• Network parameter balance (excitatory/inhibitory) is key for optimizing reservoir computing performance, not attractor dynamics.
• Systematic network design in RBNs can achieve task-specific performance at critical regimes.
• Findings guide the development of more efficient and specialized physical reservoir computers.