NeurIPS Framework Improves Brain Decoding with Anatomical Priors

A new framework called NeurIPS has been developed by researchers for decoding brain activity from fMRI data, utilizing neuro-anatomical inductive priors to address the trade-off between performance and fidelity seen in existing decoders. This innovative approach integrates a Selective ROI Spherical Tokenizer (SRST) for effective geometric encoding and a Structure-Guided Mixture of Experts (SG-MoE) that incorporates individual anatomical features. On the Natural Scenes Dataset, NeurIPS sets a new benchmark for surface-based decoders, achieving performance on par with robust 1D baselines while significantly improving convergence speed (10 epochs compared to 600). This advancement allows for quick adaptation to new subjects or tasks. The findings were published on arXiv under ID 2605.24993.