A year ago in Nature Biotechnology, Becht et al. argued that UMAP preserved global structure better than t-SNE. Now @GCLinderman and me wrote a comment saying that their results were entirely due to the different initialization choices: https://www.biorxiv.org/content/10.1101/2019.12.19.877522v1 …. Thread. (1/n)
-
-
In the single-cell context I normally run UMAP/t-SNE after reducing the data with PCA down to 50 dim. Just tried it on my laptop: on MNIST (70,000x50) FIt-SNE barely wins with 66sec vs. 72sec, but on 5x MNIST (350,000x50) FIt-SNE comfortably wins with 7min vs. 15min. [cont.]
-
[cont.] When using MNIST with all 784 dimensions ("ambient dimensions" in that figure), UMAP is indeed faster but this has nothing to do with UMAP vs t-SNE and is only due to the different kNN search implementations (annoy in FIt-SNE vs pynndescent in UMAP). [cont.]
- Još 2 druga odgovora
Novi razgovor -
-
-
Also do you agree regarding the theoretical motivations being stronger for UMAP?
-
Looking at the UMAP paper it certainly feels that way :-) but I did not form a definitive opinion about that yet.
Kraj razgovora
Novi razgovor -
Čini se da učitavanje traje već neko vrijeme.
Twitter je možda preopterećen ili ima kratkotrajnih poteškoća u radu. Pokušajte ponovno ili potražite dodatne informacije u odjeljku Status Twittera.