Necessary conditions for STDP-based pattern recognition learning in a memristive spiking neural network

Area of Science:

• Neuroscience
• Materials Science
• Computer Science

Background:

• Spiking Neural Networks (SNNs) show promise for efficient pattern recognition.
• Memristor devices offer a hardware-efficient platform for implementing SNNs.
• Unsupervised learning rules are crucial for enabling SNNs to learn from data without explicit labels.

Purpose of the Study:

• To investigate effective local training rules for unsupervised pattern recognition in memristor-based SNNs.
• To demonstrate Spike-Timing-Dependent Plasticity (STDP) for weight changes in hardware.
• To analyze learning convergence across various SNN architectures and memristive parameters.

Main Methods:

• Experimental demonstration of STDP using memristor-synapse hardware.
• Analysis of learning convergence in single-layer and complex SNNs for clusterization and digit recognition.
• Development of a probabilistic generative model for rate-based networks.
• Heuristic algorithm for experimental determination of convergence conditions.

Main Results:

• Demonstrated STDP-based weight change in a memristor-neuron system.
• Identified convergence in binary clusterization and handwritten digit recognition tasks using memristive STDP.
• Proposed the "correlation growth-anticorrelation decay" principle for optimal parameter configuration.
• Developed a heuristic algorithm to find convergence conditions, accounting for device variability.

Conclusions:

• The "correlation growth-anticorrelation decay" principle provides a near-optimal policy for configuring SNN parameters.
• Binary clusterization convergence serves as a benchmark for tuning parameters in rate-coding SNNs.
• The proposed approach is general and applicable to various memristors and SNN implementations for unsupervised learning.