We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:

Download:

Current browse context:

stat.ML

Change to browse by:

References & Citations

Bookmark

(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Computer Science > Machine Learning

Title: On UMAP's true loss function

Abstract: UMAP has supplanted t-SNE as state-of-the-art for visualizing high-dimensional datasets in many disciplines, but the reason for its success is not well understood. In this work, we investigate UMAP's sampling based optimization scheme in detail. We derive UMAP's effective loss function in closed form and find that it differs from the published one. As a consequence, we show that UMAP does not aim to reproduce its theoretically motivated high-dimensional UMAP similarities. Instead, it tries to reproduce similarities that only encode the shared $k$ nearest neighbor graph, thereby challenging the previous understanding of UMAP's effectiveness. Instead, we claim that the key to UMAP's success is its implicit balancing of attraction and repulsion resulting from negative sampling. This balancing in turn facilitates optimization via gradient descent. We corroborate our theoretical findings on toy and single cell RNA sequencing data.
Comments: 20 pages, 15 figures; minor changes, added run-times and error bars
Subjects: Machine Learning (cs.LG); Machine Learning (stat.ML)
Cite as: arXiv:2103.14608 [cs.LG]
  (or arXiv:2103.14608v2 [cs.LG] for this version)

Submission history

From: Sebastian Damrich [view email]
[v1] Fri, 26 Mar 2021 17:22:58 GMT (6499kb,D)
[v2] Thu, 22 Apr 2021 13:22:29 GMT (8354kb,D)

Link back to: arXiv, form interface, contact.