Calibrated reservoir computers

Area of Science:

• Complex systems
• Computational neuroscience
• Machine learning

Background:

• Reservoir computing is a powerful tool for time-series forecasting, especially for chaotic data.
• Understanding the performance landscape of these systems is crucial for reliable predictions.
• Existing methods struggle with the inherent complexities and uncertainties in chaotic dynamics.

Purpose of the Study:

• To investigate the fine-scaled performance alternations in reservoir computers for chaotic data forecasting.
• To understand the underlying mechanisms causing these performance variations.
• To develop a method for reducing the observed performance uncertainty.

Main Methods:

• Analysis of the performance landscape of reservoir computers.
• Investigating transversal stability variations within the synchronization manifold.
• Relating observational and internal dynamical states.

Main Results:

• Observed infinitely fine-scaled alternations in the performance landscape.
• Identified the emergence of these structures through variations in transversal stability.
• Demonstrated a link between synchronization manifold stability and performance uncertainty.

Conclusions:

• Performance uncertainty in reservoir computers for chaotic forecasting is linked to synchronization manifold stability.
• A simple calibration method can effectively attenuate performance uncertainty.
• This work provides a pathway for more reliable chaotic data forecasting using reservoir computing.