Path length statistics in reservoir computers

Area of Science:

• Computational neuroscience
• Complex systems
• Machine learning

Background:

• Reservoir computing (RC) involves high-dimensional dynamical systems, making design challenging.
• Parameter space exploration for RCs requires extensive simulation, hindering efficient design.

Purpose of the Study:

• To identify a simple statistic for predicting reservoir computer performance.
• To reduce the computational complexity of designing reservoir computers.

Main Methods:

• Analysis of the relationship between mean path length between nodes and RC performance.
• Evaluation of signal diversity using the covariance matrix of the reservoir computer's signals.

Main Results:

• A statistic based on mean path length is correlated with reservoir computer performance.
• This statistic predicts the diversity of signals generated by the reservoir, as measured by the covariance matrix.
• The statistic alone is insufficient as it doesn't account for matching to training signals.

Conclusions:

• Mean path length offers a valuable metric for initial screening of reservoir computer network configurations.
• This approach can significantly reduce the design space and computational cost by eliminating unsuitable designs early.
• Further research is needed to integrate this statistic with training signal matching for optimal reservoir computer design.