#37 · Systems biology

Video model features and generative representations predict human visual cortex fMRI…

The three preprints share a focus on predicting human visual cortex fMRI responses from video-trained or video-generative representations. The first routes V-JEPA-2 video features across space and time through a cross-attention encoding model, targeting higher visual cortex and dynamic spatiotemporal structure. The second introduces RF-DiVE and RF-GO, population receptive field-constrained most-exciting-input frameworks for encoding models spanning V1 through hV4, with in-silico validation. The third reports that representations from future-frame generation in video diffusion models align better with fMRI responses than representations of observed video. Together they extend encoding-model work toward temporally resolved, receptive-field-aware, and generative representational spaces. The common thread is the use of video models as probes of cortical visual representation; the receptive-field and generative-alignment findings are the novel elements, with the scope of each limited to the reported in-silico and fMRI settings.