Boosting Reservoir Computing with Brain-inspired Adaptive Dynamics

Area of Science:

• Computational neuroscience
• Artificial intelligence
• Machine learning

Background:

• Reservoir computers (RCs) offer efficient computation and brain-inspired principles.
• RCs simplify training by fixing internal connections but are sensitive to hyperparameters.
• Standard RCs neglect crucial excitatory/inhibitory (E-I) neuronal signal balance.

Purpose of the Study:

• Investigate the impact of E-I balance on RC performance.
• Introduce a self-adapting mechanism for E-I balance.
• Enhance RC robustness and performance across diverse tasks.

Main Methods:

• Analyzed RC performance across different E-I balance regimes.
• Developed a self-adapting mechanism to tune E/I balance locally.
• Incorporated brain-inspired heterogeneity in target neuronal firing rates.

Main Results:

• RCs perform optimally in balanced or slightly over-inhibited states.
• The self-adapting mechanism improved performance by up to 130% in memory and prediction tasks.
• Heterogeneity in firing rates reduced hyperparameter sensitivity and improved task versatility.

Conclusions:

• Dynamic adaptation of E-I balance is superior to static optimization in RCs.
• Brain-inspired mechanisms enhance RC performance, robustness, and computational understanding.
• Self-adapting RCs represent a promising direction for neural computation.