Most fMRI encoding models treat prediction as deterministic regression: a given stimulus maps to a single predicted brain response. The underlying biology, however, is not deterministic. Identical stimuli evoke a distribution of neural responses, arising from physiological noise, hemodynamic variability, and cognitive fluctuations.
MultiFlow, described in a new arXiv preprint, reframes the task. Rather than regressing to a point estimate, it applies stimulus-informed source flow matching to generate a distribution of plausible fMRI responses conditioned on the stimulus. The model is vision-driven and multimodal, integrating visual input with other stimulus information.
The significance lies in the shift from point prediction to distribution prediction. If the model captures the full spread of possible responses, it may provide a more faithful account of how the brain represents stimuli. The abstract does not yet include quantitative comparisons against deterministic baselines, so the practical gains remain to be seen.