Rastermap

Rastermap#

Clustering: n_clusters=None vs specified values#

  • n_clusters=None or 0: disables clustering, sorts individual neurons. This is only feasible for small datasets (~<200 neurons), as the sort becomes NP-hard for large N. Useful for fine-grained neuron sorting.

  • n_clusters=N (e.g. 50–200): applies scaled k-means before sorting. Each cluster is treated as a “superneuron”. Default is 100. More clusters preserve detail but slow down sorting; fewer clusters compress more but risk mixing signals.

Example:

  • 54 neurons → n_clusters=None

  • Whole-brain recording (thousands) → n_clusters=100

Clustering Step Details#

  1. PCA on (neurons × time) to extract n_PCs=200 components → spatial basis.

  2. Clustering in PCA space using scaled k-means → U_nodes (cluster centroids), X_nodes (average trace).

  3. Temporal basis comes from the cluster traces (X_nodes).

  4. If n_clusters=None: each neuron becomes its own cluster.

This step compresses data from N neurons → k clusters and denoises by averaging.

Superneurons and Binning#

  • A superneuron is the average of a group of similar neurons.

  • After sorting, neurons can be binned into superneurons for display (e.g. 50-neuron bins).

  • Binning is distinct from clustering: it’s for visualization.

  • Smaller clusters than bin size are fine; bins may cross cluster boundaries.

  • Superneurons smooth out noise and reduce display rows.

Sorting via TSP + Segment Shifting#

  • Rastermap orders clusters using a variant of the Traveling Salesman Problem (TSP).

  • Steps:

    1. Compute similarity matrix (B B^T) from cluster traces.

    2. Define a target matrix balancing global similarity and local continuity.

    3. Use heuristic segment shifting to optimize order (NP-hard problem).

    4. verbose_sorting=True shows shift steps.

Locality parameter w controls:

  • w=0: prioritize global groupings.

  • w=1: prioritize local sequences.

Time lag: allows asymmetric similarity (lead/lag in activity).

Embedding and Upsampling#

  • After cluster ordering, neurons get assigned positions in 1D.

  • grid_upsample=10 creates more refined positions between clusters.

  • Each neuron gets a fractional coordinate (embedding), not just a cluster label.

  • Enables intra-cluster ordering based on small differences.

Final Outputs#

  • embedding: fractional 1D coordinate per neuron.

  • embedding_clust: cluster ID per neuron.

  • isort: final sorted index list.

  • Sorted similarity matrix is approximately block-diagonal.

Summary: Model Step by Step#

Step

Compression

Modeling Purpose

PCA

Yes

Reduces time dimensionality, extracts signals

Clustering → superneurons

Yes

Forms prototypes, reduces problem size

TSP sorting

No

Finds 1D manifold of similarity

Segment shifting

No

Optimizes the TSP ordering

Cluster upsampling

Kind of

Fine-grains neuron placement

Neuron Binning

Yes

Smooths output for display

References#

Contents