NEVO: EPFL Videos Drive 2.3 SD Higher BOLD in Selected Voxels
NEVO inverts fMRI encoders inside video diffusion models to synthesize stimuli that maximally activate chosen brain voxels. The approach extends prior image-level neuro-optimization work and enables precise, repeatable experimental control. Clinical and regulatory pipelines will need new provenance standards for AI-generated stimuli.
The system inverts a pretrained fMRI predictor to optimize latent codes inside a video diffusion model. Training combined 1.2 million natural videos with simultaneous 7T BOLD recordings from 12 subjects. Optimization targets single voxels or small ROIs in early visual cortex, yielding stimuli whose activation maps match the target mask at r=0.81.
Related work includes the 2023 NeuroGen framework from Stanford, which optimized static images for category-selective areas, and the 2022 Mind-Video reconstruction model from Chinese Academy of Sciences that decoded 2-second clips at 4.2% top-1 accuracy. NEVO extends both by closing the loop from voxel target to generated video without intermediate semantic labels.
Operational impact appears first in stimulus design for basic neuroscience experiments. Labs can now replace hand-curated image sets with on-demand videos that isolate single-voxel tuning, cutting session time by an estimated 35%. Regulatory files for non-invasive brain stimulation devices will require disclosure of synthetic-stimulus provenance once such tools reach clinical testing.
EPFL: NEVO stimuli will appear in at least three peer-reviewed fMRI studies with n>20 human subjects by December 2025.
Sources (3)
- [1]Primary Source(https://nevo-project.epfl.ch/)
- [2]Supporting Source(https://arxiv.org/abs/2306.00923)
- [3]Supporting Source(https://www.nature.com/articles/s41562-022-01450-1)