Get notified
All projects
NeuroDataReHack 2026

Hippocampal - Prefrontal Manifold Geometry Across Novelty and Learning

Mikhail (Misha) Proskurin

Project Description

Population-geometry (“cognitive map”) analysis of hippocampal CA1 and prefrontal (PFC) ensemble recordings from the Jadhav lab, using manifold and dimensionality-reduction methods. Two W-track dandisets from the same lab are analyzed in parallel, not merged: one probing how the map differs between novel and familiar environments, the other how it evolves as an animal learns a new environment across a single day. CA1 and PFC are always processed separately.

Objectives and Approach

Questions:

  • (000447) How does the population geometry transform between novel and familiar contexts, and how do CA1 and PFC relate?
  • (000978) How does the manifold evolve session-by-session during learning, and does it change its dimensionality or just its shape?

Approach.

  • Data is streamed lazily from DANDI (pynwb/remfile). A common pipeline runs per dandiset: (1) extract time-binned spike-rate matrices with position/velocity and epoch/condition/session labels; (2) linear baselines, PCA, lap-resolved dPCA (cross-validated regularization + permutation significance), and GPFA with latents indexed by linearized track position; (3) nonlinear embeddings, UMAP and CEBRA (supervised and unsupervised CEBRA-Time) on a validated 50 ms / Gaussian-smoothed / speed-filtered representation; (4) geometry comparison, Procrustes/CCA on position-matched centroids, with tracks linearized onto a common W topology so different physical mazes are comparable; (5) intrinsic-dimensionality triangulation, TwoNN, PCA participation ratio, Isomap residual variance, and a decoding-vs-dimension curve, all cross-validated.

Progress and Next Steps

Done

  • compared in track-relative (linearized) coordinates the maps show a spatially-structured transformation, clear shared geometry plus real reshaping. dPCA confirms a genuine space×condition interaction (remap), significant in all animals.
  • 000978: the manifold converges monotonically toward its final-session geometry across the day, robust to bin size, embedding, region, and 2-D vs. track-relative binning.
  • Dimensionality: both maps are low-dimensional and curved, TwoNN/Isomap give ~3–5 intrinsic dimensions vs. a much higher linear participation ratio (the gap is a curvature signal). Intrinsic dimensionality is largely unchanged by familiarisation (000447) and, within a session, is stable (~3) across learning (000978), the pooled ~8 reflects cross-session drift, not within-session complexity. Conclusion: learning/novelty reshape the geometry of a fixed-low-dimensional map rather than changing its dimensionality.

Next Steps

  • Topology (persistent homology, ripser) to test whether the low-D geometry is the expected ring structure and whether that topology is preserved across the novel→familiar transformation.
  • Per-animal vs. pooled alignment, and consistency checks across animals for the learning trajectory.

Background and References