Three postdoctoral positions in generative AI, computational neuroscience, and neuroengineering.

The problem

Treatment selection in psychiatry is fundamentally a counterfactual problem: what would happen to this patient under a different treatment?

We want to build generative world models that learn how brain activity, behavior, facial expression, language, symptoms, and treatment evolve together, and use those models to estimate individualized treatment trajectories and counterfactuals.

What you could work on

Develop new methods for multimodal generative modeling, diffusion and flow models, causal and counterfactual inference, representation learning, and behavioral modeling; train them on longitudinal clinical datasets spanning brain imaging and multimodal behavioral data; and develop methods for testing when patient-specific counterfactual predictions are identifiable, reliable, and clinically useful.

We are also particularly interested in candidates working on cross-species pose estimation and behavioral representation learning, including methods that identify shared and species-specific structure in movement, behavior, and latent neural or affective states across humans and animal models.

This is a methods-development position, not simply an application of existing models to clinical data. The exact direction can be shaped around the fellow's interests.

The problem

Can we learn a patient-specific dynamical model of the brain well enough to predict, and ultimately control, its response to stimulation?

We are developing generative digital twins of neural dynamics from TMS and high-density EEG, with the long-term goal of moving from prediction to individualized, closed-loop neurostimulation.

What you could work on

Develop models of patient-specific neural dynamics and intervention response; solve forward and inverse problems linking stimulation to distributed brain activity; develop methods for system identification, causal dynamics, generative modeling, and model-based control; and test these models using clinical TMS-EEG data from patients undergoing treatment.

The goal is both methodological and translational: new approaches to learning controllable neural dynamics, tested in a setting where the intervention can actually be delivered.

The problem

Many of the hardest measurement problems in neuroscience share the same computational structure: we observe high-dimensional, indirect measurements and want to infer the underlying biological sources.

We are interested in whether neural operators, differentiable physical models, and modern inverse methods can make previously intractable inverse problems fast enough, scalable enough, and accurate enough to solve at experimental scale.

What you could work on

Develop new methods for large-scale inverse problems and operator learning, with projects spanning computational microscopy, multi-camera imaging, neural source localization, and TMS-EEG.

Potential directions include learning mappings between optical measurements and three-dimensional neural activity; computational imaging through complex or scattering optical systems; fast reconstruction for volumetric microscopy; multi-view and multi-camera reconstruction; EEG source localization; estimating distributed neural responses to TMS; and learning reusable forward and inverse operators that generalize across subjects, geometries, acquisition systems, or experimental conditions.

Methodologically, this could involve neural operators, physics-informed and physics-constrained learning, differentiable simulation, learned iterative reconstruction, inverse-problem regularization, uncertainty quantification, and hybrid approaches that combine mechanistic forward models with learned components.

The fellow could focus primarily on computational imaging, on electromagnetic and source-localization problems, or develop methods that span both. The common goal is to develop general computational approaches for recovering latent biological structure from large, ill-posed measurement systems.