We develop new computational and theoretical methods to reveal fundamental principles of brain function based on brain activity signals. Four ongoing directions — each anchored in real neuroimaging or electrophysiology data.
Models that jointly explain electrophysiology, neuroimaging, genetic, and behavioral data — from single-neuron biophysics to whole-brain effective connectivity.
The brain operates across many time and length scales at once: ion-channel dynamics on the millisecond scale, hemodynamic responses on the second scale, and behavior on the minute scale. Our group builds models that bridge these scales explicitly, with data-constrained parameters.
A central tool is dynamic causal modeling (DCM) extended to partially observed multimodal data. We have shown how metabolic, thalamo-cortical, and hierarchical DCM variants can be estimated from EEG and fMRI recorded during rest, disease, or pharmacological challenge — including clozapine-treated schizophrenia and pediatric epilepsy.
Ongoing work extends these models to ketogenic modulation of pediatric epilepsy and to inferring absence-seizure mechanisms in childhood absence epilepsy through thalamo-cortical DCM.
Activity-dependent and homeostatic plasticity models designed to inform personalized brain-treatment schemes — modulating activity toward healthy attractors.
We formulate optimal-control problems on neural systems endowed with realistic plasticity rules. In prior work, we described a computational framework for controlling the self-restorative brain based on free-energy and degeneracy principles, and a framework for optimal control of self-adjustive neural systems with activity-dependent and homeostatic plasticity.
Current directions include TMS-induced brain-dynamics modeling and personalized grid-based whole-brain response modeling to characterize how repetitive single-pulse TMS reshapes networks.
End-to-end EEG/fMRI pipelines — from preprocessing and connectivity estimation to brain-state decoding and behavioral prediction — with machine learning built in and every step scripted for reproducibility.
We treat analysis as software: structured configurations, version-controlled code, and SLURM-based batch processing on HPC clusters make every result reproducible — from raw EEG or fMRI data to final figures. Standard preprocessing (SPM, FieldTrip) feeds into connectivity estimation and into state-based descriptions of brain dynamics.
On top of these pipelines we apply machine learning to neural signals: transformer-based deep learning for EEG seizure detection, MDS/MDS+ and energy-landscape embeddings of brain-state dynamics, and supervised decoding that links brain states to behavior or clinical outcomes — always with rigorous cross-validation.
Entropy-based measures, maximum-entropy models, and AI-driven optimization — quantifying how brain networks store, process, and transfer information.
Information theory offers model-free ways to ask what neural signals tell us. We develop and systematically evaluate entropy-based directionality estimators for multivariate time series, and estimate pairwise maximum entropy models (pMEM) of nonlinear brain-state dynamics with empirical Bayes — yielding energy landscapes that expose how resting-state networks transition between cognitive states.
The same combinatorial problems — optimal network partitioning, community detection, and homophily structure — are also tackled with AI-based optimization: quantum and quantum-inspired approaches such as QAOA alongside classical graph and ensemble methods.