A new AI architecture borrows a trick from the hippocampus to decide which of its internal experts should handle a task.
Researchers built SpikeMoE, a framework that merges spiking neural networks with mixture-of-experts models. The key piece is a new router modeled on competition-inhibition dynamics seen in the brain's CA1 region: it uses lateral inhibition and a refractory period to let experts compete via spike counts, then selects the top performers for a given input. The team also added a module for handling missing sensory inputs in multimodal tasks, blending prototypes from available data with learnable stand-in embeddings. Tested on vision, language, and multimodal benchmarks, SpikeMoE outperformed other spiking neural network baselines and matched or beat conventional neural networks, while staying robust when chunks of input data were missing.
Spiking neural networks fire in discrete pulses rather than continuous values, which makes them a natural fit for low-power neuromorphic chips, but they have historically trailed standard neural networks on accuracy. Pairing that efficiency with mixture-of-experts sparsity, which only activates a subset of a model's parameters per task, is a bet that two efficiency tricks compound rather than cancel out. Closing that accuracy gap while keeping the energy profile intact is the real headline here, not the brain-inspired framing.
The efficiency case still rests on specialized neuromorphic hardware that most data centers don't run, so the energy savings are closer to a lab promise than a deployed one for now.