A Stretchable Leaky Integrate-and-Fire Neuron Enables Uncertainty Handling in Spiking Neural Networks

Flexible neuromorphic computing systems have emerged as a promising candidate for next-generation edge computing systems, especially for the development of lightweight, low-power spiking neural network (SNN) systems. However, the mechanical coupling noise caused by flexible components poses challenges in uncertainty handling during the inference process in SNNs. Herein, we developed a stretchable leaky integrate-and-fire neuron device with enhanced uncertainty handling capabilities for stochastic SNN hardware implementation. These neuron devices utilize strain-induced uncertainty to model neuronal function as a conditional random variable. In addition, the SNN was incorporated in a Bayesian approximation model when the firing patterns exhibit mean stationarity over time steps. Simulations on classification and regression tasks on various datasets performed improved confidence and high generalization capability. The results demonstrate that the stretchable LIF neurons have significant potential for constructing robust and accurate SNNs for flexible neuromorphic computing.