Stochastic binary synapses having sigmoidal cumulative distribution functions for unsupervised learning with spike...

Area of Science:

• Neuromorphic Engineering
• Computational Neuroscience
• Materials Science

Background:

• Spike Timing-Dependent Plasticity (STDP) is crucial for synaptic plasticity in neuromorphic systems.
• Continuous weight control in STDP poses hardware implementation challenges.
• Previous simplified STDP with binary synapses suffered from memory degradation during learning.

Purpose of the Study:

• To propose a novel stochastic binary synaptic model for simplified STDP.
• To improve memory maintenance and recognition accuracy in unsupervised online learning.
• To enable hardware implementation using memristor-based switching devices.

Main Methods:

• Developed a stochastic binary synaptic model with sigmoidal probability evolution for weight changes.
• Implemented the model using serially connected binary memristors in switching devices.
• Simulated unsupervised learning of MNIST images using a two-layer network.

Main Results:

• The proposed model demonstrated superior memory maintenance compared to conventional continuous weights.
• Achieved 97.3% recognition accuracy on MNIST, outperforming standard STDP.
• The learning rule showed robustness against device-to-device variability in memristor behavior.

Conclusions:

• The novel stochastic binary synaptic model enhances simplified STDP for neuromorphic hardware.
• This approach offers improved performance and hardware feasibility for unsupervised learning.
• The model's resilience to variability makes it suitable for practical memristor-based applications.