ideas + tasks

[sa-lee]

From meeting with @dicook yesterday

In order of priority

• explore some embeddings on the datasets mentioned in UMAP paper
• make a shiny (or vega) interface to explore all of these things
• add in variable cross filter (see Ursula's example in functionSectioning)
• come up with some criterion for evaluating the fit based on the simplex/centroids approach
• explore other tour variants especially cmass as this is conceptually similar to t-sne's use of a gaussian kernel
• trying t-sne in 3-d, then tourring the tetrahedron that will form from a triangulation here
• ask @uschiLaa for particle physics dataset - evaluate the use of t-SNE as run on that data

[sa-lee]

So far have uploaded the mnist and fashion mnist data to this repo. Next step is to provide the interface for exploring the embeddings interactively.

[sa-lee]

Also need to let the tours go for longer period of times... are there heuristics around this for a given dataset size?

[sa-lee]

most of these ideas have gone into the design spec on the wiki