Simple deterministically constructed cycle reservoirs with regular jumps.

Area of Science:

• Computational neuroscience
• Machine learning
• Time series analysis

Background:

• Reservoir computing, a class of state-space models, utilizes a fixed reservoir with an adaptable readout for time series processing.
• Echo state networks (ESNs) are a prominent, yet typically randomized, type of reservoir model.

Purpose of the Study:

• Introduce and evaluate a novel deterministic reservoir model, the cycle reservoir with jumps (CRJ).
• Investigate the relationship between reservoir characteristics and model performance.
• Develop a framework to precisely measure the short-term memory capacity of linear reservoir models.

Main Methods:

• Designed a novel deterministic reservoir model (CRJ) with constrained weights.
• Analyzed reservoir properties like eigenvalue distribution and pseudo-Lyapunov exponents.
• Introduced a new framework for assessing short-term memory capacity.

Main Results:

• CRJ models consistently outperformed standard ESNs across diverse temporal tasks.
• CRJ models achieved superior performance despite a more constrained eigenvalue distribution than ESNs.
• Optimal CRJ models exhibited consistently negative pseudo-Lyapunov exponents, unlike ESNs.

Conclusions:

• The CRJ model presents a powerful and efficient alternative to standard ESNs for time series modeling.
• Reservoir dynamics and eigenvalue properties are crucial for understanding model performance.
• Shortcut connections in CRJ topology significantly impact memory capacity, offering avenues for further optimization.