Large-scale neural recordings are usually analyzed one dataset at a time. A model is trained for a particular experiment, a particular set of neurons, and a particular downstream task. This produces useful local solutions, but it limits what can be learned across the rapidly growing collection of neural datasets.

My research asks a broader question: what structure in neural population activity is reusable?

Transfer is more than predictive accuracy

A transferable neural model should support several forms of reuse. It should adapt to a new recording with limited target-specific parameters, support multiple downstream tasks, and preserve a representation space in which independently collected datasets remain comparable.

These goals are related but not identical. A model may forecast a new dataset well after full retraining while losing all cross-dataset comparability. Conversely, a frozen backbone may achieve similar predictive performance while offering a shared coordinate system and a more efficient adaptation protocol.

Why population dynamics?

Neural computation is expressed through coordinated activity over time. Modeling this activity as a population process creates an interface between machine learning and systems neuroscience: forecasting tests whether temporal structure has been captured; decoding tests whether task variables are accessible; representation analysis asks whether the learned space reflects functional or anatomical organization.

A practical research program

  • Learn from large, heterogeneous collections of recordings.
  • Adapt with small modules rather than retraining the entire model.
  • Evaluate across sessions, animals, laboratories, experimental paradigms, and species.
  • Connect predictive performance to interpretable biological structure.
The long-term goal is not a single universal model that erases experimental differences, but a reusable modeling language that makes those differences easier to compare.

This note is part of an evolving research notebook. Publication claims should be cited from the linked papers rather than from this informal summary.