A Hybrid CMOS-Memristor Spiking Neural Network Supporting Multiple Learning Rules.

Area of Science:

• Neuroscience and Computer Engineering
• Artificial Intelligence and Neuromorphic Computing

Background:

• Artificial intelligence (AI) faces challenges with traditional computing, particularly the von Neumann bottleneck, due to high memory access demands.
• Memristive devices offer a solution for low-latency, energy-efficient in-memory computing by emulating synaptic plasticity.
• Brain-inspired spiking neural networks (SNNs) leverage synaptic plasticity for enhanced computational capabilities.

Purpose of the Study:

• To develop a neuro-synaptic architecture integrating two distinct learning rules on a single memristor-based platform.
• To demonstrate the co-integration of spike-timing-dependent plasticity (STDP) and Bienenstock-Cooper-Munro (BCM) learning rules.
• To validate the architecture's effectiveness in unsupervised learning tasks.

Main Methods:

• Electrical characterization and modeling of memristor devices.
• Design of a novel neuro-synaptic architecture for co-integrating synaptic devices.
• Implementation of STDP and BCM learning rules within the proposed architecture.
• Testing the architecture on two distinct unsupervised learning tasks.

Main Results:

• Successful co-integration of STDP and BCM learning rules using a single type of synaptic memristor device.
• Demonstration of a unique platform for implementing multiple plasticity rules.
• Effective performance of the architecture in addressing two different unsupervised learning challenges.

Conclusions:

• The proposed neuro-synaptic architecture effectively utilizes memristive devices to implement diverse learning rules.
• This approach offers a pathway towards more powerful and efficient brain-inspired computing systems.
• The co-integration of plasticity rules enhances computational capabilities for AI applications.