An Online Structural Plasticity Rule for Generating Better Reservoirs

Area of Science:

• Computational Neuroscience
• Machine Learning
• Neuromorphic Engineering

Background:

• Liquid state machines (LSMs) utilize recurrent neural networks of spiking neurons for temporal processing.
• Training LSMs, particularly the 'liquid' component, remains a challenge, often relying on random or less efficient methods.
• Biological neural networks exhibit structural plasticity, a mechanism for adapting connections.

Purpose of the Study:

• To propose a novel neuro-inspired unsupervised learning rule for training the reservoir of LSMs.
• To mimic biological structural plasticity by dynamically forming and eliminating synaptic connections.
• To evaluate the performance improvements in separation properties, classification accuracy, and memory retention.

Main Methods:

• Developed a low-resolution, online unsupervised learning rule inspired by structural plasticity.
• Implemented the rule to dynamically rewire synaptic connections within the reservoir.
• Utilized Address Event Representation (AER) protocols for updating connection matrices in memory.
• Assessed performance through pairwise separation, linear separation, generalization ability, and fading memory analysis.

Main Results:

• Trained liquids demonstrated 1.36x greater interclass separation and 2.05x better linear separation than random liquids.
• The proposed rule maintained similar intraclass separation and generalization abilities compared to random liquids.
• Trained liquids showed a longer fading memory (83.67 ms) compared to random liquids (92.8 ms) for specific inputs.
• Outperformed a separation-driven synaptic modification algorithm in liquid separation and classification accuracy across multiple pattern recognition tasks.

Conclusions:

• The novel neuro-inspired learning rule effectively trains liquid state machines by dynamically rewiring synaptic connections.
• This approach significantly enhances key performance metrics, including data separation and classification accuracy.
• The method offers a biologically plausible and efficient alternative for training neuromorphic systems.