Growing Echo-State Network With Multiple Subreservoirs.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Recurrent Neural Networks

Background:

• Echo-State Networks (ESNs) offer an alternative to gradient-based training for recurrent neural networks (RNNs).
• Determining the optimal reservoir structure for ESNs remains a significant challenge for specific applications.

Purpose of the Study:

• To introduce a Growing Echo-State Network (GESN) for automatic reservoir size and topology design.
• To address the limitations of fixed-structure ESNs in adapting to diverse applications.

Main Methods:

• Utilizes block matrix theory to incrementally add hidden units, forming multiple subreservoirs.
• Employs predefined singular value spectra for subreservoir weight matrices to ensure the echo-state property.
• Features incremental updates of output weights during network growth and provides convergence proof.

Main Results:

• The GESN demonstrated superior prediction performance on artificial and real-world time-series benchmarks.
• Achieved faster learning speeds compared to ESNs with static reservoir configurations.
• Successfully validated the automatic design of reservoir size and topology.

Conclusions:

• The proposed GESN effectively automates reservoir design for ESNs.
• GESNs offer improved performance and efficiency over traditional fixed-structure ESNs.
• This method provides a robust solution for applying ESNs to various time-series problems.