Statistically chosen to maximize *some particular loss measure, which in this case might be the t-SNE or UMAP criterion, and is computed only globally and not for different filters.
Right (I mean, I'm saying "right" but really I should just say "I'm taking your word for it"), but even more fundamentally this is dimensionality reduction from an OpenAI embedding vector, which seems almost like the asymptotic limit of inscrutability.
