What you're looking at. Every model turns a card into a long list of numbers.
UMAP
squashes those down to two so they can be drawn:
cards that sit near each
other are cards the model thinks are alike. One panel per model, same cards in
every panel.
What to trust. Nearness is meaningful; the distance
between two blobs is
not, and neither is which way up a panel is. Panels are rotated onto the first one so
they're comparable, but the real answer to "do these models agree?" is the
neighbour-agreement table at the bottom, measured on the full-size vectors rather
than the picture.
How to use it. Click a dot to see the card and what each model puts next to it;
hover those names to preview them. Search highlights matches in every panel at once.
Each panel's
?
explains what that model was fed. Built with
umap-learn.
Tap a dot for the card · two fingers to move and pinch to zoom · one finger scrolls the page