#Umap
UMAP 0.6 is out! A new initialization, new optimizers and faster nearest neighbour search make UMAP much faster and far better at large scale. On 3 million arXiv papers it finishes in under 5 minutes, before UMAP 0.5 has even found its nearest neighbours.
November 11, 2026 at 1:27 PM
Last week, I got a first glance at pose embeddings from a subset of my poetry slam corpus!

During a two-day workshop on multimodal embeddings, we got to play around with the new embedding exploration tool by @tillmannohm.bsky.social: Eidora

@dhssfau.bsky.social
@mila-oiva.bsky.social
November 16, 2026 at 8:36 AM
🎙️Podcast: CHINA TAGEBUCH
🎤Folge 287: ALLE VERRUECKT

Es geht um Bücher bzw. die deutsche Buchbubble, DE‑Presse über CN, um eine wilde Räuberpistole und natürlich um Jazz.

Nano
Elektroflug
Taiwan
Nvidia
uMap
Shanghai Jazz
LinkTree
Bücher

🎤: diary.umlauts.de?podcast=287-...
🎥: youtu.be/ls3wjCJBF-I
November 17, 2026 at 7:42 AM
The code is all available on GitHub:

github.com/lmcinnes/umap

and you can install via PyPI or conda-forge.

$ pip install -U umap-learn
GitHub - lmcinnes/umap: Uniform Manifold Approximation and Projection
Uniform Manifold Approximation and Projection. Contribute to lmcinnes/umap development by creating an account on GitHub.
github.com
November 11, 2026 at 1:27 PM
A new UMAP release brings major performance, quality, and scaling improvements
Discussion | hackernews | Author: lmcinnes

#MachineLearning
A new UMAP release brings major performance, quality, and scaling improvements
umap-learn.readthedocs.io
November 15, 2026 at 5:39 AM
I guess the next bottleneck is the frontend and having tools to easily select through those millions of datapoints that we can now easily UMAP 😅
UMAP 0.6 is out! A new initialization, new optimizers and faster nearest neighbour search make UMAP much faster and far better at large scale. On 3 million arXiv papers it finishes in under 5 minutes, before UMAP 0.5 has even found its nearest neighbours.
November 12, 2026 at 8:15 AM
Documentation on the options available to better handle large million-scale datasets is also available:

umap-learn.readthedocs.io/en/latest/la...
Embedding Very Large Datasets — umap 0.6.0 documentation
umap-learn.readthedocs.io
November 11, 2026 at 1:27 PM
Can confirm, the new UMAP rocks. I usually work with embeddings between 10k and 100k, and the speedup is as advertised. Important for me as I work currently with restricted CPU resources.
Neighbour search is ~4x faster by making better use of PyNNDescent. Initialization is faster with a new init method. And using the new optimizers, the whole run is reproducible with a fixed seed while using every core. A seeded 0.5 run is single-threaded: 52 minutes, 12x longer.
November 12, 2026 at 6:34 AM
🎙️Podcast: CHINA TAGEBUCH
🎤Folge 287: ALLE VERRUECKT

Es geht um Bücher bzw. die deutsche Buchbubble, DE‑Presse über CN, um eine wilde Räuberpistole und natürlich um Jazz.

00:52 Schreibmonat
03:10 Elektroflug
06:01 Taiwan - Xi
09:50 Jensen Huang
14:42 uMap […]

[Original post on literatur.social]
November 17, 2026 at 7:42 AM
UMAP was originally designed around a small portion of the attraction-repulsion spectrum (www.jmlr.org/papers/v23/2...). The new UMAP 0.6 optimizers lift these restrictions, allowing UMAP to perform well for much higher repulsion strengths, providing better kNN recall.
November 11, 2026 at 1:27 PM
Join us on Monday, Dec. 14th at 11 am ET for a SUPeR Viewer workshop - we'll be going over how to create your own SUPeR view that can be easily incorporated into the BAR's single cell data portal for plants. Register here: forms.gle/TnBSCeGKDRoq...
SUPeR Viewer workshop registration, 14-Dec-2026
Join us on Monday, December 14th at 11 am ET for a 1 hour introduction on creating a new view for the BAR's single cell Single Cell UMAP/Pseudospatial eFP BrowseR (SUPeR) Viewer based on your own sing...
forms.gle
November 17, 2026 at 7:09 PM
Currently the new optimizers and initialization are opt-in as they change the output: set compatibility_layout=False. Future versions will make them the default, so try them now! Documentation on the new optimizers is available here:

umap-learn.readthedocs.io/en/latest/op...
Choosing a UMAP Layout Optimizer — umap 0.6.0 documentation
umap-learn.readthedocs.io
November 11, 2026 at 1:27 PM
A large part of the overall improvement comes from a brand-new initialization scheme. The new recursive initialization works by initializing UMAP using UMAP itself -- laying out a coarse version of the graph structure, and then refining, in a recursive tower of initializations.
November 11, 2026 at 1:27 PM
“Stop Misusing t-SNE and UMAP for Visual Analytics”, by Hyeon Jeon et al. #ieeevis ieeevis.org/year/2026/pr...
November 10, 2026 at 11:25 PM
The qualitative examples against UMAP with default parameters look very compelling. Could you explain how the examples compare with parameterized UMAP?
November 12, 2026 at 3:32 PM
It is worth comparing the increased detail and fidelity provided by the new UMAP 0.6 optimizers. They resolve structure at much finer scales, and can significantly improve on the neighbourhood accuracy of UMAP 0.5 despite being faster.
November 11, 2026 at 1:27 PM
Neighbour search is ~4x faster by making better use of PyNNDescent. Initialization is faster with a new init method. And using the new optimizers, the whole run is reproducible with a fixed seed while using every core. A seeded 0.5 run is single-threaded: 52 minutes, 12x longer.
November 11, 2026 at 1:27 PM
And sure, this is a not so subtle subskeet against a family of analytics tools that throw a SPLOM over here, a stream graph over there, and a umap projection in the middle.

I always find myself a bit overwhelmed when looking at those UIs
November 13, 2026 at 5:51 PM
📊 Tutorial #10: Data Visualization Fundamentals

Choose plots for the question, check preprocessing, and interpret patterns carefully. Examples use synthetic data.

theomicshub.com/tutorials/da...

#Bioinformatics #DataViz
Data Visualization Fundamentals
Choose and interpret QC plots, PCA, UMAP, heatmaps, and volcano plots with explicit preprocessing and cautious biological conclusions.
theomicshub.com
November 17, 2026 at 7:31 AM
The new recursive initialization provides a far better starting point, much closer to UMAP's final result than the messy blurs provided by spectral initialization or PCA, and it does so faster than either!
November 11, 2026 at 1:27 PM
Just as important as how quickly we get a result is the quality of the result. UMAP 0.6 works far better on million-scale datasets thanks to a new, better initialization and significant improvements in the optimizer. The result resolves finer cluster structure in fewer epochs.
November 11, 2026 at 1:27 PM
New CRAN package manifoldsR with initial version 0.5.0
#rstats
https://cran.r-project.org/package=manifoldsR
CRAN: Package manifoldsR
R interface to the Rust crates 'manifolds-rs', 'ann-search-rs' and 'evoc-rs'. Builds 2D (or higher) embeddings with UMAP (McInnes et al., 2018) &lt;<a href="https://doi.org/10.48550%2FarXiv.1802.03426" target="_top">doi:10.48550/arXiv.1802.03426</a>&gt;, t-SNE (van der Maaten and Hinton, 2008) &lt;<a href="https://jmlr.org/papers/v9/vandermaaten08a.html" target="_top">https://jmlr.org/papers/v9/vandermaaten08a.html</a>&gt;, densMAP and den-SNE (Narayan et al., 2021) &lt;<a href="https://doi.org/10.1038%2Fs41587-020-00801-7" target="_top">doi:10.1038/s41587-020-00801-7</a>&gt;, PHATE (Moon et al., 2019) &lt;<a href="https://doi.org/10.1038%2Fs41587-019-0336-3" target="_top">doi:10.1038/s41587-019-0336-3</a>&gt;, PaCMAP (Wang et al., 2021) &lt;<a href="https://jmlr.org/papers/v22/20-1061.html" target="_top">https://jmlr.org/papers/v22/20-1061.html</a>&gt;, ForceAtlas2 (Jacomy et al., 2014) &lt;<a href="https://doi.org/10.1371%2Fjournal.pone.0098679" target="_top">doi:10.1371/journal.pone.0098679</a>&gt;, DREAMS and diffusion maps. Also generates approximate nearest neighbour graphs that every method can reuse, plus EVoC and k-means clustering. Multi-threaded via 'Rayon', with f32 or f64 precision.
cran.r-project.org
November 17, 2026 at 1:02 PM
📄 "Stop Misusing t-SNE and UMAP for Visual Analytics" by Hyeon Jeon, Jeongin Park, Sungbok Shin, and Jinwook Seo. #ieeevis

Learn more at https://ieeevis.org/year/2026/program/content/8f097b32-b19a-4015-852e-65c2b68adc89/
November 10, 2026 at 7:15 PM
Most impressive project!

Would UMAP be applicable to something like the Bluesky Atlas?
bsky.app/profile/jaz....
jaz.sh Jaz @jaz.sh · Sep 27
Okay so I've built a new Atlas, this time it lets you explore the conversations on Bluesky over the past 7 days graphically.

It should update every ~6 hours and help you navigate the information environment that exists on here.

I've learned a LOT about what people do here today...

atlas.jazco.dev
Atlas
A living map of what Bluesky is talking about: the past week's conversations, grouped into topics and regions and rebuilt every six hours.
atlas.jazco.dev
November 11, 2026 at 4:52 PM