Evolving Dual-Threshold Bienenstock-Cooper-Munro Learning Rules in Echo State Networks.

Area of Science:

• Computational Neuroscience
• Machine Learning

Background:

• Existing Bienenstock-Cooper-Munro (BCM) learning rules face limitations with synaptic long-term depression (LTD) affecting low-activity synapses, potentially causing information loss.
• Biological experiments suggest a dual LTD threshold mechanism that allows for potentiation, depression, or no change even in activated synapses.
• Current BCM rules utilize fixed parameters, lacking biological plausibility and flexibility for learning complex input signal structures.

Purpose of the Study:

• To propose an evolved dual-threshold BCM learning rule to optimize reservoir internal connection weights in echo-state networks (ESNs).
• To address information loss and enhance learning performance in ESNs by introducing adaptable LTD thresholds.
• To investigate the synergistic learning capabilities of the proposed rule.

Main Methods:

• Development of an evolved dual-threshold BCM learning rule.
• Application of the rule to regulate reservoir internal connection weights in ESNs.
• Experimental evaluation on benchmark tasks and esterification process prediction.

Main Results:

• The evolved dual-threshold BCM learning rule effectively alleviates information loss.
• The proposed rule enhances the overall learning performance of ESNs.
• Synergistic learning of different plasticity rules was observed, outperforming existing methods.

Conclusions:

• The evolved dual-threshold BCM learning rule offers a more biologically plausible and flexible approach to synaptic plasticity in ESNs.
• This novel rule demonstrates significant improvements in learning performance and information processing capabilities.
• The findings suggest broader applicability in complex signal processing and prediction tasks.