SpikingMoE brings Mixture-of-Experts to Spiking Neural Networks for energy-efficient AI

Spiking Neural Networks (SNNs) promise ultra-low power consumption for AI, but they often struggle with dynamic task adaptation. SpikingMoE, a new framework from researchers at arXiv (Yang et al., May 2026), tackles this by marrying SNNs with the Mixture-of-Experts (MoE) paradigm. Unlike traditional dense networks, MoE activates only a subset of specialized “expert” modules per input, reducing compute cost. SpikingMoE goes further by making the entire pipeline spike-driven—each expert uses binary spike communication, and routing is controlled by a spike-driven prompt (SDprompt) modeled after the brain's lateral geniculate nucleus (LGN). This biologically plausible design allows dynamic expert fusion without continuous analog signals, making it directly compatible with neuromorphic chips like Intel Loihi.

On standard benchmarks, SpikingMoE demonstrates practical viability: it hits 94.09% top-1 accuracy on CIFAR-10 and 74.54% on CIFAR-100, numbers competitive with much larger, energy-hungry models. This proves that modular, input-dependent computation can be achieved in SNNs without catastrophic accuracy loss. As the first open-source SNN framework to integrate MoE with LGN-inspired routing, SpikingMoE opens the door for energy-efficient, adaptive visual recognition in edge devices—think always-on cameras, drones, or smart sensors that can reroute their internal processing on the fly. The paper's release on arXiv positions it as a foundational resource for researchers exploring neuromorphic MoE systems.