Spiking neural networks beat ANNs in fMRI visual decoding by 27x

A new study on arXiv (2607.19170) investigates spiking neural networks (SNNs) as alternative feature extractors for fMRI-based visual semantic decoding. Traditional methods use features from artificial neural networks (ANNs) as regression targets, but it's unclear whether these align well with actual brain activity. The authors compare one ANN baseline against four SNN variants from the same architectural family, differing only in spiking dynamics. To isolate feature quality, all models use the same L2-regularized linear decoder—only the target feature vectors change.

Results on the GoD dataset are striking: SNN-derived features reduce feature-prediction error from 0.7707 (ANN) to 0.0282, and improve top-1 semantic decoding accuracy from 18% to 44%. Ablation studies confirm both spiking neural dynamics and the number of temporal simulation steps contribute to the gain. These findings indicate SNN features are more brain-decoded and highlight target feature design as a critical component for next-generation brain-computer interfaces and neural decoding systems.