Podcast with Allard Roubroeks on neuroimaging and fmri analysis

How do we move beyond "this brain region lights up" to genuinely understanding how neural circuits compute? Allard Roebroek argues that the future lies in merging bottom-up computational models with top-down neuroimaging analysis , and that neither community can succeed alone. Subscribe for more from the Convergent Science Network podcast series. Allard Roebroek tackles a fundamental tension in neuroimaging: the field generates gigabytes of whole-brain data per minute, yet most analyses reduce this richness to statements about which regions activate during which tasks. He distinguishes two modeling traditions that have developed largely in isolation. Bottom-up modelers build biophysically inspired simulations of neural circuits, from spiking networks to hemodynamic coupling, but face a crippling indeterminacy problem: infinitely many models can reproduce the same behavioral data. Top-down modelers invert observation models to go from fMRI or EEG signals back to inferred neural activity, but typically work with only a handful of pre-selected brain regions and stop at causal connectivity without asking what computations those regions perform. Roebroek's vision is to unite these approaches. He advocates for models that simultaneously perform the task (as bottom-up models do), are biophysically grounded, and are accountable to whole-brain neuroimaging data (as top-down models aspire to be). This triple requirement has not yet been achieved, but he argues it is the only path toward models that are both computationally meaningful and empirically constrained. The whole-brain coverage of fMRI provides a unique advantage over electrophysiology , not in spatial or temporal resolution, but in the ability to observe the entire system at once without invasive intervention. The interview engages seriously with criticisms of this program. Can correlation-based neuroimaging data really constrain causal models? Roebroek acknowledges that causality requires assumptions beyond correlation, but argues that computational models themselves provide exactly those assumptions , transforming observed dependencies into mechanistic explanations. He also confronts the common practice of restricting analyses to regions of interest, which discards the very whole-brain information that makes neuroimaging valuable, and calls for models that encompass at least all cortical regions plausibly involved in a given task.