SpikemoE integrates brain-like routing with energy-efficient ai models
SpikeMoE: Brain-Inspired Competitive Routing for Flexible Spiking Mixture-of-Experts
Artificial Intelligence
Summary
Spiking Neural Networks mimic how brain neurons send signals using spikes, which helps computers process information efficiently. The authors combined this idea with a method that picks the most relevant small networks, or experts, to work on a problem, inspired by how parts of the brain compete to be active. They built a system called SpikeMoE that uses this brain-like competition to select which experts to use, especially helpful when some types of input data are missing. Their tests show that SpikeMoE works well on vision, language, and tasks involving multiple senses while saving energy compared to traditional methods.
What this means in practice
- •For embedded system engineers: Design energy-efficient AI models that can flexibly select specialized components based on spiking neuron activity for resource-constrained devices.
- •For multimodal application developers: Build robust AI systems that handle incomplete sensory inputs by generating missing data representations using a combination of prototype and learned embeddings.
Authors
Xiaoli Liu, Yujie Liang, Jialin Li, Malu Zhang
Abstract
Spiking Neural Networks (SNNs) enable event-driven computation through biologically inspired dynamics at the neuronal scale, while Mixture-of-Experts (MoE) perform conditional computation through expert selection at the model scale. Integrating their strengths offers potential for flexible neural architectures. A key challenge, however, lies in designing an expert selection mechanism based on spiking activity. To address this, we introduce a spike-based k-WTA Router inspired by competition-inhibition observed in the hippocampal CA1 region. The router incorporates lateral inhibition and refractory period to select Top-K experts according to discrete spike counts. Building on this, we present SpikeMoE, a framework that integrates neuronal-scale spiking dynamics with model-scale expert selection. To address incomplete multisensory inputs in multimodal tasks, we further equip SpikeMoE with a two-stage missing-modality modeling module that combines empirical prototypes from an observed-modality pool with modality-specific learnable embeddings to construct missing-modality representations. Experiments on vision, language, and multimodal benchmarks demonstrate that SpikeMoE achieves state-of-the-art performance among the SNN baselines, matches or exceeds the performance of ANN counterparts, and maintains robustness across diverse missing-modality conditions. These results demonstrate a favorable trade-off between performance and energy efficiency, validating the integration of spiking dynamics with sparse expert computation and highlighting SpikeMoE as a promising approach to energy-efficient brain-inspired computing.