Boosting reservoir computing with brain-inspired adaptive control of E-I balance

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Reservoir computers (RCs) are recurrent neural networks inspired by brain principles, offering an efficient alternative to deep learning.
• RCs typically use fixed internal connections and trained output weights, but are sensitive to hyperparameters.
• The excitatory-inhibitory (E-I) signal balance, crucial for brain function, is usually fixed in RCs, despite its impact on performance.

Purpose of the Study:

• To investigate the impact of tuning the E-I signal balance in RCs.
• To introduce and evaluate a self-adapting mechanism for E-I balance adjustment.
• To explore the potential for improved RC performance and reduced hyperparameter tuning.

Main Methods:

• Systematic investigation of varying E-I balance ratios in RCs.
• Development and implementation of a self-adapting mechanism to locally adjust E-I balance based on target firing rates.
• Performance evaluation using memory capacity and time-series prediction tasks, including experiments with heterogeneous firing-rate targets.

Main Results:

• Optimal RC performance was consistently observed in balanced or slightly over-inhibited E-I regimes, not excitation-dominated ones.
• The self-adapting mechanism significantly reduced hyperparameter tuning costs.
• Performance gains of up to 130% were achieved in memory capacity and time-series prediction tasks.
• Incorporating heterogeneity in firing-rate targets further improved robustness.

Conclusions:

• Dynamic adaptation of the E-I balance is a key design principle for enhancing RC performance.
• The proposed self-adapting mechanism offers a practical approach to optimize RCs and reduce tuning complexity.
• Findings provide insights into neural computation and suggest avenues for developing more efficient and robust artificial neural networks.