Evolving robust memristive spiking neural networks via metaheuristic algorithms for noise-resilient neuromorphic inference

Abstract

Neuromorphic systems hold the promise of ultra-low-power event-conditions intelligence, but noise and non-idealities of hardware of the memristive devices limit their practical use. Nonetheless, it is not understood how to attain an inference which is both noise-robust and accurate as well as energy-efficient across realistic multi-source perturbations. In this case we develop a metaheuristic evolutionary learning system which substantially alleviates this well-known robustness–efficiency trade-off in memristive spiking neural networks. The optimization of spiking networks is done with a noise-aware form of evolutionary strategy that explicitly takes into consideration event-based input noise, spike-timing jitter, conductance variability and long-term drift. Several noise levels are tested and compared with a three-level gradient trained spiking network, ANN-to-SNN conversion and non-evolved memristive SNN with fixed parameters in the context of the task of gesture recognition based on events. At low noise (ρ = 0.1) all four models reach similar accuracies of around 95% (n = 336). The proposed MSNN shows 87–91 percent accuracy at moderate noise (0.40–0.6), while the accuracy of gradient-based SNNs and ANN-to-SNN models are 70–80 percent and below 60 percent, respectively. Even with extreme noise (0.8–1.0), the accuracy is 76–82 compared to 51% and ANN-to-SNN models break to approximately 25%. Normalized robustness analysis shows that the highest loss of relative accuracy is about 0.21 which is three to four times less than baseline models. These findings indicate that evolutionary optimization reveals inaccessible noise tolerant representations with sparse and temporally structured spiking. This framework lays the groundwork to feasible, energy-efficient and resilient memristive neuromorphic systems and offers the new structure of applying SNNs to noisy hardware systems.