DNA + image
Without CL
View DNA, image and label embeddings trained together without the contrastive loss (CL).
Open supplementary demo →Explore embedding training
Play through training, inspect individual samples, and explore taxonomic cones in a native 2D embedding space.
Primary demo · without image
DNA-only
Start here. Follow DNA embeddings and taxonomic labels through training. This is the experiment used for the project’s 2D method illustration.
Open DNA-only demo →Supplementary demos
Additional image-based experiments, provided for exploration.
DNA + image
DNA + image
CL weight 0.1
View joint DNA and image training with the contrastive loss weighted by 0.1.
Open supplementary demo →Image-only
View image embeddings and taxonomic labels without the DNA branch.
Open supplementary demo →On iPad: tap a point to inspect it, drag with one finger to pan, and pinch or use + / − to zoom. Use Play or drag the timeline to move through training. Scroll the page outside the plot.
Each demo covers five orders, 575 labels and 197 snapshots from step 0 to 19,545, with 216 fixed training probes per enabled modality. These are native 2D subset experiments, not projections of the 768D result models or held-out evaluations. Cone checks indicate whether a probe lies inside its true label cone; they are separate from nearest-label prediction.
Viewer pages are approximately 27–38 MB each and load only when opened.