The steps are to start by identifying the clique of 4 nodes with the largest sum of weight, and to build new faces (triangles) from there. The algorithm naturally searches within a cluster before finding the exit. By construction the graph is always connected. No need for the full correlation matrix
extra: I went into more details on some of the motivation which led to working on this project last year.
bsky.app/profile/spii...
If you work at the intersection of network science, manifold learning, or want to map massive tabular datasets into GNNs, this provides a highly scalable way to build sparse, parsimonious graphs.
Preprint available on arXiv! arxiv.org/abs/2603.09564 5/5
By restricting the algorithm's "memory" to a sliding window of the exploration frontier, we drastically cut compute bloat. We can process 100k nodes in ~500 seconds: a fraction of the time exact methods take for just 25k nodes. 4/5
A 2D UMAP projection of a 100,000 node graph showing distinct, color-coded, branching fractal-like structures instead of solid blobs.
Here is an a-TMFG for 100,000 nodes generated from a Gaussian Markov Random Field. Instead of dense "blobs", it preserves the beautiful, dendritic branching structure and intrinsic cluster boundaries of exact TMFGs. 3/5
Standard TMFG construction requires a dense correlation matrix. Exact methods hit a wall around ~25k nodes. We bypassed this using a k-NN initialization, HNSW indexing, and a bounded face universe with lazy-deletion which resulted in near-linear scaling! 2/5
From a weekend side project to pre-print:
How do you learn a sparse graph which encodes short range interactions? Triangulated Maximally Filtered Graphs (TMFGs) are perfect for this, but they scale terribly (O(N²)).
Excited to share my new paper introducing a-TMFG which resolves the issue! 1/5
It's this shows weakness despite great animation.
It was funny at the beginning but it's increasingly overdone.
I welcome it. We were promised a bunch of new features which have yet to materialize. Jay has done a great job launching the platform but clearly the lack of dynamism over the past year, dare I say, the inability to capitalize on momentum shows something, or maybe someone else, is missing
You will have other opportunities given this is "only" the beginning of this presidential term. I remember taking advantage of price swings on $TLSA when Trump and Musk were publicly bickering. Anticipating/reacting to Trump is a full time job unfortunately.
New acquisition
I was invited by @eliesgf.bsky.social to give an online seminar to @jenseisert.bsky.social's q. machine learning group. I spoke about "Fundamentals of q. Boltzmann machine learning w/ visible & hidden units" arxiv.org/abs/2512.19819. Slides from my talk available here:
zenodo.org/records/1889...
jesus, it's such a terrible show with great animation.
NEW: The bad news about international-student enrollments at American colleges just got worse. An exclusive @chronicle.com analysis of just-released State Department data shows new visa issuances in the summer of 2025 dropped by more than a third. www.chronicle.com/article/the-...
How convenient
Just seeing this gem.
Are AI models effective collaborators, or mere assistants awaiting your next command? (Preprint: arxiv.org/abs/2602.24188)
To find out, we make AI collaborate with itself, in private information games: tasks that require sharing private information, like this chess board ordering task.
Undergraduate student interested in molecular dynamics, quantum chemistry, or ML for materials?
Join the @cecamevents.bsky.social Summer School at SISSA (Trieste), June 22–July 3, 2026. Designed specifically for beginners.
Details & apply:
www.cecam.org/workshop-det...
The economist trying to save the planet with his own crystal ball: read @doynefarmer.bsky.social's interview in The Standard
#ComplexityEconomics #EnergyTransition @smithschool.ox.ac.uk @inetoxford.bsky.social
www.standard.co.uk/lifestyle/ec...
marble qr code: ah, I just had a silly idea, dynamic qr code. Since this lattice was recorded at the critical temperature. You could embed an ordered lattice in all this noise. Nobody asked for new qr codes I know 🫠
Once again, a long post with strong opinions. It's probably twice as long as it should be, it's also repetitive and written in affect. And you probably disagree with my argument. So maybe you shouldn't read it. On the other hand, most things worth reading are written in affect.
It's clearly been traumatized by the concept of the void
The 2D Potts model on a square lattice near the critical temperature.
Just a clarification: is that one of those open claw bots?
Bro got snubbed by the Nobel peace committee and he's been hijacking, and bombing left and right ever since. He's stopped counting how many wars he's "stopped". Really makes you think.
Yes, having fun working through a couple visualizations on old projects.
bsky.app/profile/ocra...
