Cloud Distance Lab
Two Gaussian point clouds, A and B, and eight ways to say how far apart they are. Drag a cloud or its center, or reshape it with the sliders, and see where the Sinkhorn distance used by CogniMune agrees with the exact earth mover’s distance and six other measures — and where they part ways.
Pick a scenario to see where the distances agree and where they part ways.
Projection on PC1
Separation sweep
Show sweep data as a table
How the lab relates to CogniMune
Each point stands in for a cell and each cloud for one blood sample, drawn in just two marker dimensions. The 40-color panel describes every cell with 39 markers, but the question is the same: how far apart are two distributions of cells?
Whole distributions, not just centers
The distance between centroids is blind to spread, shape and orientation — try “Same center, wider B”. Transport distances account for where every point sits.
Distances that keep growing
Histogram measures saturate once the clouds stop overlapping — try “No overlap”: Jensen–Shannon caps at 1 bit and Bhattacharyya becomes infinite, while transport distances keep measuring the gap.
Sinkhorn trades exactness for speed
Entropic regularization cuts the cost of optimal transport (from roughly O(n³ log n) to ~O(n²) per iteration, per the proposal). As ε shrinks, the Sinkhorn value approaches the exact earth mover’s distance — compare their two cards.
In the lab the ground cost is the Euclidean distance, so the Sinkhorn and exact EMD values share the plot's units. The CogniMune analysis engine computes Sinkhorn distances between immune-cell population distributions to build the graph fingerprints described on the Technology page. All points here are synthetic and generated in your browser; they are not study data.
One dimension, step by step
The live Sinkhorn demo shows the transport plan between two histograms and how it sharpens as ε shrinks.