A compound memristive synapse model for statistical learning through STDP in spiking neural networks.

Area of Science:

• Neuromorphic Engineering
• Computational Neuroscience
• Materials Science

Background:

• Memristors are promising for mimicking biological synapses in neuromorphic computing.
• Fabricating reliable nanoscale memristive synapses with continuous conductance changes is challenging.
• Existing approaches face difficulties in achieving reliable synaptic function.

Purpose of the Study:

• Propose a novel compound memristive synapse design using binary memristors.
• Investigate the computational implications of synaptic plasticity in this new design.
• Explore its potential for unsupervised learning in spiking neural networks.

Main Methods:

• Developed a compound memristive synapse using parallel bistable memristors.
• Integrated stochastic filament formation into an abstract model of stochastic switching.
• Analyzed spike-timing dependent plasticity (STDP) and unsupervised learning in winner-take-all networks.

Main Results:

• Compound memristive synapses exhibit STDP with stabilizing weight dependence.
• The network implements generalized Expectation-Maximization for unsupervised learning.
• Networks accurately represent high-dimensional data distributions (e.g., handwritten digits) under noise and device variations.

Conclusions:

• Compound memristive synapses offer a viable alternative to continuous-state memristors.
• This design principle facilitates robust unsupervised learning in neuromorphic systems.
• The approach holds promise for future neuromorphic architectures requiring high fidelity and resilience.