Leland McInnes
@lelandmcinnes.bsky.social
A Mathematician dabbling in Data Science, especially unsupervised learning and data exploration. UMAP, HDBSCAN, PyNNDescent, DataMapPlot. (He/Him)
Agents are everywhere now but is worth remembering that there are still so many cool things we can do with good ol matrix factorization and embeddings. No LLMs required.
David Gasquez@davidgasquez.com · 2mo agoHow community notes reduce viral misinformation. Great talk/conversation about community notes internals and impact. www.ted.com/talks/keith_...
Wikimedia concepts embedded and then mapped to a sphere with UMAP. Trying out topic clusters as continents. Not really sure where it's going.
New in the Archives Browser (Members): Semantic Space — CAN archives as a 3D cloud. Every post embedded, mapped with UMAP, clustered with k-means. Hover for titles, click to trace relatives + search. #threejs → www.creativeapplications.net/archives-bro...
The Shape of Attraction in UMAP: Exploring the Embedding Forces in Dimensionality Reduction Mohammad Tariqul Islam, Jason W. Fleischer Action editor: Dmitry Kobak https://openreview.net/forum?id=fdPNhqav5G #embedding #cluster #dimensionality
Want to compare your molecules to FDA-approved drugs or the Enamine 10k library? I made an open-source tool that visualizes physicochemical properties and metrics like Böttcher complexity as 2D/3D PCA or UMAP plots. Fast and easy. Mac only. github.com/sculpturatus...
Need to run your latent space exploration inside a Jupyter Notebook instead? No worries! We have got you covered there as well! (5/n)
🔭 Paper Day! 🧪 Introducing Hyrax — an astro-focused ML framework built for the era of Rubin, Roman, and Euclid Handles the entire ML lifecycle: from data acquisition → training → inference → interactive latent-space discovery. pip install hyrax (1/n) 🧵 #astronomy #astrocoding #extragalactic
Can neural representations from different people be translated into a shared space without paired data? In our preprint, "Platonic Representations in the Human Brain: Unsupervised Recovery of Universal Geometry", with Rishi Jha and @fuentemilla.bsky.social, we test this using fMRI 🧵👇
![Marcos-Manchón, P., Jha, R., & Fuentemilla, L. (2026). Platonic Representations in the Human Brain: Unsupervised Recovery of Universal Geometry. arXiv [q-Bio.NC]. Retrieved from http://arxiv.org/abs/2605.20496](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:dlr4kkceoy6mwbryfqlnryd7/bafkreicl7tyf3fowavfsim4pbdzxgqbk53glyjnhprwb5twwrqwzo7ijwq)
Showcase #2: 🔭 Cluster-Scale Lens hunting with Hyrax in collab with @khullar.bsky.social UMAP + DBSCAN found a galaxy cluster overdensity in @vrubinobs.bsky.social DP1 → human-in-the-loop spotted a lensed arc behind it → Euclid VIS confirmed presence Gemini follow-up incoming 👀 (9/n)
New post: Tagging my blog with BERTopic and LLMs. I suspect as token costs increase, we'll see more blended traditional ML/LLM systems. And, in general, they work really well! vickiboykis.com/2026/05/18/t...
Tagging my blog posts with BERTopic and LLMs
LLMs mean you still need a human in the loop, but in a different part
vickiboykis.com
High-dimensional data is hard to understand. But is it truly cursed? To help you build better intuition for high-dimensional datasets, @lekschas.de and I developed dtour, a visualization tool for smoothly navigating through data projections.
Seth Levine gave a great talk on the art of clustering at ODSC East that really digs in the practical realities of getting the most out of your data. www.youtube.com/watch?v=LtZ3...
The Art of Clustering: The Good, The Bad and The Beautiful by Seth Levine, Contentsquare
YouTube video by Open Data Science and AI Conference
youtube.com
People should not sleep on the new EVoC clustering algorithm. It's got me rethinking a lot of my widgets and bulk labelling tricks. youtu.be/1roN51JXuX8
marimo
The Most Satisfying Method to Explore Embeddings
youtube.com
What’s the world searching for right now? Explore the past 24 hours in combined google trends and local news sources from 125 countries. Similar topics cluster together. Focus on the clusters or the outliers. 🐸 👉 trends.datafrosch.fun #dataviz #dataexploration worldnews #UMAP #localmodels
EVoC is a library designed specifically for fast clustering of high dimensional embedding vectors. It can produce high quality clusters extremely efficiently, and requires little to no hyperparameter tuning. Better clustering than UMAP + HDBSCAN; faster clustering than KMeans.
There's a new release of fast-hdbscan out now on PyPI that adds some major new features. - A PLSCAN implementation based on work by Jelmer Bot - Support for diverse metrics via PyNNDescent - A powerful "cannot-link" constraint system for semi-supervised clustering by Richard Hakim
One of the nicest things about Nvidia model releases is that they ship the training data. What does it look like? I sampled 250k examples from 24 datasets in the Nemotron post-training v3 collection and built an interactive Embedding Atlas to explore it.
This week I'll be delivering a talk and leading a tutorial for the Language AI in Space Sciences Workshop in Baltimore, hosted at STScI. So excited to see folks from all disciplines join the conversation -- including astronomers, engineers, computer scientists, librarians, and linguists.
Han Jiao (ex jina Ai ceo) built mlx apple silicon clustering libs for umap, tsne etc github.com/hanxiao/mlx-...
Excited to share a new #interactive #dataviz, putting 50 years of #neuroscience research on the map: The State of Neuroscience 2025 stateofneuroscience.thetransmitter.org #StateOfNeuro
State of Neuroscience 2025: Trends & Breakthroughs | The Transmitter
A comprehensive look at major trends shaping the neuroscience landscape in 2025
stateofneuroscience.thetransmitter.org
Any sufficiently large k-nn is indistinguishable from magic 🧙♂️
gavin leech@gleech.org · 5mo agoNew paper on a long-shot I've been obsessed with for a year: How much are AI reasoning gains confounded by expanding the training corpus 10000x? How much LLM performance is down to "shallow" generalisation (approximate pattern-matching to highly-related training data)? t.co/CH2vP0Y7OF
I made a map of 3.4 million Bluesky users - see if you can find yourself! bluesky-map.theo.io I've seen some similar projects, but IMO this seems to better capture some of the fine-grained detail
Bluesky Map
Interactive map of 3.4 million Bluesky users, visualised by their follower pattern.
bluesky-map.theo.io
Our new pre-print shows how unsupervised clustering methods can identify biologically meaningful differences in early vocal production, with no human feedback. @antorrisi.bsky.social has led this interdisciplinary collaboration based on computational methods + #chicks 🐣 arxiv.org/abs/2601.12203
here's a fun side project i've been working on: i compiled a joint text<>audio embedding model to a fast coreml pipeline, and built a very fast (~400ms for 50k samples, can scale to millions) UMAP dimensionality reduction GPU impl in mlx. using it to browse music libraries and do sample sim search
