This is where everything converges. Can we use large language models to understand the brain - and brain data to build better AI? The BLUM paper tests whether LLMs that make human-like speech errors also show human-like lesion patterns. We "lesion" LLMs (ablating layers, perturbing weights) and map their error profiles onto real brain anatomy using Symptom-Lesion Models trained on 213 stroke patients. If the mapping holds, it means LLM architectures are recapitulating something about how the brain organises language - not just mimicking outputs.

At allt.ai, we're building tools that go the other direction: applying what we've learned from 20 years of neuroimaging to make AI systems that handle language more like brains do. Meanwhile, our deep learning morphometry work uses CNNs to predict aphasia severity directly from brain structure (Communications Medicine, 2024), and our machine-learning multimodal models combine MRI, demographics, and lesion data to forecast language recovery outcomes.

Tools we've contributed to: Brainchop (edge AI for neuroimaging in the browser), NiiVue (web-based neuroimaging visualization), and dcm2niix/ezBIDS (cloud-friendly DICOM conversion).

Through the Aging Brain Cohort I built a multimodal lifespan MRI database linking brain structure, genetics, cognition, and behaviour in healthy adults. The core finding: brain age - the gap between your brain's apparent age and your chronological age - predicts real outcomes. Advanced brain age predicts worse aphasia (Neurology, 2023), and we've shown it mediates the link between cardiovascular health and language recovery (Aging Brain, 2025).

White matter hyperintensities are the connective tissue of this thread (literally). We've published a series showing that WMH load mediates age-cognition relationships, tracks with socioeconomic status, HbA1c levels, pulse pressure, and small vessel disease - even independent of stroke lesions. At SC-ADRC, we connect this to Alzheimer's risk via APOE4, FOXP2 gene expression, and periodontal health markers.

As part of C-STAR ($23M+ NIH), I've worked on understanding why some people recover language after stroke and others don't. We released the Aphasia Recovery Cohort as an open-source repository (Scientific Data, 2024), and the Stroke Outcome Optimization Project for acute ischemic strokes. Our lesion-symptom mapping work spans traditional approaches through to gene-expression-based methods (FOXP2 mapping in Journal of Neuroscience, 2026).

Key findings: progressive lesion necrosis in chronic stroke predicts worsening aphasia; cerebellar atrophy - far from the lesion site - independently predicts language processing deficits; and cortical hierarchy (the sensorimotor-association gradient) modulates how lesions affect language networks.

My doctoral and postdoctoral work established that the human mirror neuron system is more active during complementary actions than imitative ones (Nature Neuroscience, 2007). This was a surprise - the prevailing theory was that mirror neurons are for imitation. Instead, they're for cooperation: understanding what your partner needs so you can do your part.

This led to a series of fMRI studies on joint action, cooperative bar-lifting, empathic error monitoring, and action prediction. We showed that dorsal pars opercularis is critical for both biological and non-biological action cues, and that self-identification modulates error-related brain activity when watching friends vs. rivals. The thread also included second-language acquisition (my PhD thesis) and the neural basis of gestural learning.

I directed the USC Brain Stimulation Laboratory for over a decade, using tDCS and TMS to modulate brain function across domains: motor skill acquisition in the non-dominant hand, inhibitory control in nicotine addiction (theta-burst to IFG), and rehabilitation in stroke and TBI. We also used stimulation as a probe - virtual TMS lesions of pars opercularis confirmed its causal role in action understanding.