Research
What I study
I'm interested in how the brain holds a decision in mind and acts on it — and in the quieter problem underneath it: who are the neurons doing the computing? My work moves between large-scale recordings in behaving primates and the applied-math tools needed to make sense of them.
Decision-making
How prefrontal cortex commits to a choice
I train rhesus macaques on perceptual and context-dependent decision-making tasks and record from dorsolateral prefrontal cortex with laminar and Neuropixels probes — hundreds of sessions of extracellular activity while the animal weighs evidence, holds it in working memory, and selects an action.
Rather than reading single cells one at a time, I treat the population as a dynamical system and ask about its computational geometry: how stimulus, memory, and motor signals are arranged in neural state space, and how that arrangement supports flexible behavior. To test mechanism, I model the same tasks with low-rank recurrent neural networks and compare them to the data.
Population activity as trajectories through neural state space (illustrative).
Cell types
Reading cell types from a spike's shape
Extracellular electrophysiology is powerful but nearly blind to cell identity: every neuron is just a voltage blip. With WaveMAP, I showed that nonlinear dimensionality reduction of action-potential waveforms recovers a rich diversity of putative cell types — structure that lines up with firing properties and, increasingly, with molecular ground truth (eLife, 2021; STAR Protocols, 2023).
Since then I've pushed toward multimodal identification — composing waveform, firing, and anatomical signals into a single graph-based view of cell-type space (Nature Communications, 2026) and learning cross-modal representations with contrastive methods (ICLR Spotlight, 2025). The goal: turn a spike train back into a labeled, interpretable population.
Waveforms embedded into putative cell-type clusters, WaveMAP-style (illustrative).
Methods & tools
An applied-math toolkit for neural data
Threaded through everything is a methodological interest, rooted in my applied-math training: graph-based and nonlinear dimensionality reduction, multimodal data integration, contrastive representation learning, and low-rank RNNs as interpretable models of computation.
I care about making these usable by others — releasing protocols and open tools so a method doesn't end at the figure. My earlier work at the Allen Institute on the mouse visual cortex (2-photon and intrinsic-signal imaging) still shapes how I think about standardized, large-scale physiology.
Low-rank recurrent networks as testable models of computation (illustrative).