Quaternion-valued echo state networks.

Area of Science:

• * Computational neuroscience
• * Signal processing
• * Machine learning

Background:

• * Existing echo state networks (ESNs) are limited in processing high-dimensional data.
• * Recent advancements in quaternion nonlinear activation functions enable new network architectures.
• * Processing 3-D/4-D data requires methods that can handle complex signal properties.

Purpose of the Study:

• * Introduce Quaternion-valued Echo State Networks (QESNs) for 3-D and 4-D processes.
• * Develop second-order optimal QESNs for both circular and noncircular quaternion signals.
• * Enhance the analysis of complex data in fields like renewable energy and human-centered computing.

Main Methods:

• * Development of QESNs utilizing novel quaternion nonlinear activation functions.
• * Implementation of widely linear QESNs using augmented quaternion statistics.
• * Modification of the standard widely linear model for dynamical reservoir properties.

Main Results:

• * QESNs effectively process 3-D/4-D data, including wind modeling and inertial body sensors.
• * Widely linear QESNs achieve second-order optimality by exploiting covariance and pseudocovariance.
• * Simulations demonstrate accurate prediction on benchmark and real-world noncircular data.

Conclusions:

• * QESNs offer a powerful framework for analyzing complex multi-dimensional time-series data.
• * The widely linear approach in QESNs rigorously accounts for signal noncircularity.
• * This work advances the application of recurrent neural networks in complex dynamic systems.