Stochastic sampling via synaptic delay in spiking RBMs using integrated resistive and threshold switching devices.

Area of Science:

• Neuromorphic Engineering
• Artificial Intelligence
• Materials Science

Background:

• Spiking neural networks (SNNs) are promising for low-power neuromorphic hardware due to their spatiotemporal processing capabilities.
• Spiking restricted Boltzmann machines (spiking RBMs) require neuron and synapse stochasticity for stable learning and inference.
• Limited intrinsic randomness in SNNs, especially with uniform inputs, can hinder learning performance.

Purpose of the Study:

• To present a delay-based hardware strategy for introducing stochasticity in spiking RBMs.
• To utilize synaptic propagation delay as a source of randomness for sampling.
• To develop an efficient hardware primitive for stochasticity in neuromorphic systems.

Main Methods:

• A synaptic unit cell integrating a synapse for weight storage and a delay module for temporal stochasticity was designed.
• The delay module employs serially integrated resistive random-access memory (RRAM) and threshold-switching (TS) devices.
• Compact modeling and circuit-level simulations were used to verify CMOS compatibility and delay behavior.

Main Results:

• Synaptic propagation delay was demonstrated to serve as a source of stochasticity for spiking RBMs.
• The RRAM resistance effectively tuned the TS turn-on delay, with delays following a log-normal distribution.
• MNIST learning using spiking RBMs with delay-based stochasticity achieved higher accuracy than baseline and conventional methods.

Conclusions:

• The RRAM-TS-based synaptic delay circuit provides an efficient hardware primitive for introducing stochasticity into neuromorphic systems.
• This method avoids the need for complex and power-consuming external random number generators.
• The proposed strategy enhances learning performance in SNNs without compromising efficiency.