Network structure effects in reservoir computers

Area of Science:

• Complex Systems
• Computational Neuroscience
• Machine Learning

Background:

• Reservoir computers are nonlinear dynamical systems applied to tasks like signal prediction and robotic control.
• They typically involve a large network of interconnected nonlinear nodes driven by input signals.

Purpose of the Study:

• To investigate the effect of altering network structure on reservoir computer performance.
• To characterize reservoir networks by edge manipulation and symmetry analysis.

Main Methods:

• Constructing reservoir networks with +1 or 0 edges and modifying them by flipping edges to -1.
• Using the fraction of flipped edges and network symmetries to characterize structural changes.
• Analyzing the rank of the covariance matrix of node time series.

Main Results:

• Flipping edges in the reservoir network alters its structure.
• The number of flipped edges influences the rank of the covariance matrix of node time series.
• Network symmetries are useful for characterizing network structure.

Conclusions:

• The rank of the covariance matrix is a key factor in understanding reservoir computer performance.
• Structural modifications, specifically edge flipping, provide a means to tune reservoir computing capabilities.