Finding hidden structure in complex systems.

We apply unsupervised multimodal learning to structural MRI to identify a candidate dissociation pattern between structural and functional neurodegeneration in Alzheimer's disease. Presented at the Workshop on Subtle Visual Computing, CVPR 2026, Denver, and published in the proceedings.

Standard evaluation of learning-based MRI reconstruction focuses on random undersampling, leaving open how reconstruction fails under physically realistic perturbations. Using a k-space, physics-grounded framework on fastMRI, we show corruption structure determines the qualitative character of failure independently of severity: structured line dropout produces catastrophic collapse (SSIM drop 0.877) while random undersampling degrades gradually (0.134) — failure modes aggregate PSNR cannot distinguish.

Using the THINGS and THINGSplus object-concept databases, we show that arousal ratings are uniquely resistant to prediction from other perceptual and semantic dimensions — evidence that arousal captures a distinct and poorly understood axis of human object perception.

A model can be accurate for the wrong reasons. Drawing on the epistemological gap between true and justified belief, we formalize the Justification Deficit (JD) — a diagnostic measuring how far a model's confident predictions rest on causally irrelevant features. Across two benchmarks with known causal structure, high accuracy does not guarantee low JD, and standard training routinely trades epistemic legitimacy for predictive performance. PhilML workshop (poster), ICML 2026.

An interpretable multimodal clustering pipeline on the OASIS-1 cohort (384 participants) integrates ICV-normalized cortical volumetrics with cognitive assessments (MMSE, CDR) to recover five reproducible phenotypic partitions (adjusted Rand index ≥ 0.80). One candidate resilience subgroup (n = 35) shows temporal–parietal atrophy comparable to impaired groups yet preserves cognition (mean MMSE 28.4) — a structural–functional dissociation unexplained by age or education, positioning multimodal phenotyping as a hypothesis-generation stage for systems-immunology research. ASI workshop, ACM-BCB 2026.

Goodhart's Law is usually read as proxy degradation under optimization. We argue a prior structural condition matters more: every evaluation framework is defined over a bounded ontological space, so entire classes of failure lie outside it by construction and stay invisible. We formalize this as Ontological Closure and propose a four-part framework for evaluation and governance under open-ended deployment, grounded in clinical-AI uncertainty and out-of-distribution perception for autonomous driving. EIML workshop, ICML 2026.

Two years of wet-lab work at Northwestern's Klein Lab investigating nNOS inhibition as a therapeutic strategy against AβO-induced neurotoxicity and TDP-43 mislocalization in MC65 and HT22 cell lines. Manuscript under peer review at the Journal of Neuroscience.