Neural representational dissimilarity quantifies how response distributions differ across stimuli, brain areas, tasks, or models, and is fundamental for comparing neural codes. Existing distance metrics rely on diverse assumptions and require separate estimation procedures. This paper shows that many of these metrics can be expressed within a flow‑matching framework originally developed for deep generative models, manifesting as Jeffreys divergences under distinct velocity constraints $$D_J(p\|q)=D_{KL}(p\|q)+D_{KL}(q\|p)$$. Empirical results demonstrate that flow matching offers superior accuracy for distributions that are complex or continuous, and the framework naturally supports the principled design of new metrics. In sum, flow matching provides a unified approach for understanding, estimating, and creating neural representational dissimilarity measures.
Review: By bridging flow‑matching techniques from generative modeling to neuroscience, the study improves both computational efficiency and flexibility in metric design, marking a valuable cross‑disciplinary contribution.