What we study, and what we build to study it.

Most detailed evidence about brain connections comes from tracer studies in macaques and rodents. In humans, diffusion tractography estimates pathways indirectly, and an estimate can look right while routing a connection through the wrong tissue. We calibrate the human estimates against ground truth: tract tracing and optical microscopy in non-human primate brains that are also imaged with MRI, so the same fibers are seen at every resolution. The result we are after is a validated, population-level wiring diagram of the human brain, with uncertainty attached, usable in models of perception and cognition and in surgical planning.

In more than 1,700 brains from age 2 to 88, regions of early visual cortex that represent the same part of the visual field are more strongly connected than regions that do not, and the connections are stronger along the horizontal than the vertical meridian and in the lower than the upper field, matching known differences in perception. From these data we are deriving projection rules: which pathways preserve retinotopic order and which mix it. We are now testing the rules at the axon level with macaque tracing, light-sheet imaging, and human post-mortem tissue, to separate real asymmetries from artifacts of tractography.

Learning to read or to calculate recruits visual, language, memory, and control systems at once, so several pathways may change together. We compare the pathways that link the regions a skill uses against control pathways, and use the timing of change to separate acquisition from retention, across a school year, an intervention, or the summer break. A baseline scan predicts individual learning outcomes; dense longitudinal imaging turns that single prediction into a trajectory. The same design serves questions about white-matter lateralization in aging, the development of memory systems, and recovery after traumatic brain injury.

Our platform records every analysis as a graph linking data, application versions, parameters, and outputs, so a result can be re-executed or re-examined years later. We develop data standards and the tools that get raw scanner data into them, and we run community benchmarks that measure how methods generalize; one shows that deep-learning tract segmentation trained on young adults fails on children. The next step is agents that turn a research question into computations across archives while recording every dataset, tool, and decision they used, so that a result produced with AI can be audited like one produced by hand.

Across the jurisdictions we have surveyed, transfer of neural data is almost always permitted in principle but conditional in practice, with conditions that differ by direction, and legal protection becomes uncertain once data are turned into features, embeddings, or trained models. The burden falls hardest on researchers in low- and middle-income countries. Our work makes the rules usable: classifying a planned transfer by risk, choosing a lawful pathway, and expressing the obligations in a form that platforms and AI agents can enforce, so that a model can be trained across archives without the data leaving them.