The following chart shows a sample of several hundred recently registered visual marks from the PTO database.
-
They are arranged so that marks with similar semantic content tend to appear near each-other, and marks with semantically distinct content will be more spaced out.
-
The underlying latent space has ~50k dimentions. Proximity in this chart is a rough proxy for proximity in that space. More complex notions of similarity are possible.
-
Mousing over an image will show the full-size image and a text description. The text descriptions is computed exclusively from the image's position in the latent space, and uses no other information about the image. The text description is also limited to ~10 words. If I computed a longer description it might be more accurate or detailed.