Unsupervised Learning on Resistive Memory Array Based Spiking Neural Networks.

Area of Science:

• Neuromorphic Engineering
• Materials Science

Background:

• Spiking Neural Networks (SNNs) show promise for efficient computing due to biological plausibility.
• Resistive Random Access Memory (RRAM) is a key emerging technology for neuromorphic hardware implementation.
• Variability in RRAM devices poses challenges for practical SNN deployment.

Purpose of the Study:

• To propose a novel training approach for SNNs compatible with RRAM hardware.
• To address the inherent conductance variations in RRAM devices.
• To enhance the feasibility of RRAM-based neuromorphic systems for online training.

Main Methods:

• Developed a greedy training approach for SNNs, focusing on temporal spike dilution and input encoding.
• Utilized Spike-Timing-Dependent Plasticity (STDP) as an unsupervised learning rule.
• Demonstrated STDP on one-transistor-one-resistor (1T1R) RRAM devices with designed voltage pulses.
• Simulated handwritten digit recognition on the MNIST dataset.

Main Results:

• The proposed training method mitigates the need for numerous RRAM conductance levels.
• Trained SNNs exhibit immunity to cycle-to-cycle and device-to-device RRAM variations.
• The approach improves cooperation between SNNs and non-ideal RRAM devices.

Conclusions:

• The greedy training method enhances the robustness of SNNs on RRAM hardware.
• This approach offers high feasibility for RRAM array-based neuromorphic systems capable of online training.
• The findings pave the way for more practical and efficient neuromorphic computing solutions.