Shubhendu Trivedi
@shubhendu.bsky.social
Interests on bsky: ML research, applied math, and general mathematical and engineering miscellany. Also: Uncertainty, symmetry in ML, reliable deployment; applications in LLMs, computational chemistry/physics, and healthcare.
https://shubhendu-trivedi.org
https://shubhendu-trivedi.org
Reposted by Shubhendu Trivedi
Wasserstein geometry = quotient geometry of permutation invariance.
In this blog, I explain why this is the natural language for exchangeable particles—and why mean-field neural network training shows up as a W2 gradient flow.
mufan-li.github.io/OT2/
In this blog, I explain why this is the natural language for exchangeable particles—and why mean-field neural network training shows up as a W2 gradient flow.
mufan-li.github.io/OT2/
Everyone Should Learn Optimal Transport, Part 2
In the previous blog post, we saw that optimal transport gives us calculus on the space of probability distributions. In this post, we will continue the core message, but we will also see that Wassers...
mufan-li.github.io
February 14, 2026 at 9:36 PM
Wasserstein geometry = quotient geometry of permutation invariance.
In this blog, I explain why this is the natural language for exchangeable particles—and why mean-field neural network training shows up as a W2 gradient flow.
mufan-li.github.io/OT2/
In this blog, I explain why this is the natural language for exchangeable particles—and why mean-field neural network training shows up as a W2 gradient flow.
mufan-li.github.io/OT2/
Reposted by Shubhendu Trivedi
"Today is ours". Mr. Padamsi on the Gujaratis of Calicut.
www.youtube.com/watch?v=UcFi...
www.youtube.com/watch?v=UcFi...
600 Years in Kerala | The Gujarati Community of Calicut | Documentary | afterImages
YouTube video by afterImages
www.youtube.com
February 10, 2026 at 11:46 PM
"Today is ours". Mr. Padamsi on the Gujaratis of Calicut.
www.youtube.com/watch?v=UcFi...
www.youtube.com/watch?v=UcFi...
Where is Claude code for packing up a few thousand books when you need it?
February 11, 2026 at 12:35 AM
Where is Claude code for packing up a few thousand books when you need it?
I was curious if Obsidian was an actual Armenian surname. Turns out not originally, but after some corruptions and misspellings of some mysterious original name, it exists as an ultra rare surname.
Anything you can do in Obsidian you can do from the command line.
Obsidian CLI is now available in 1.12 (early access).
Obsidian CLI is now available in 1.12 (early access).
February 11, 2026 at 12:25 AM
I was curious if Obsidian was an actual Armenian surname. Turns out not originally, but after some corruptions and misspellings of some mysterious original name, it exists as an ultra rare surname.
Since bibliographies were/are in the conversation recently, a random tip on bib health: If you cite an arXiv paper with copy paste from some default channels, which renders in references as <title>, <year>, <arxiv url>, it will very rarely get picked up by google scholar.
My reaction to hallucinated references in LLM chats is that they are kind of funny, in as much as they often involve papers which I would like to exist, and often by authors who could very credibly write them. There's also a slight aspect of fanfic to them that I can sort of get behind.
February 10, 2026 at 4:50 PM
Since bibliographies were/are in the conversation recently, a random tip on bib health: If you cite an arXiv paper with copy paste from some default channels, which renders in references as <title>, <year>, <arxiv url>, it will very rarely get picked up by google scholar.
Reposted by Shubhendu Trivedi
Prof. P. P. Divakaran (1936-2025) passed away last year. An appreciation by those who knew him.
bhavana.org.in/puthan-puray...
bhavana.org.in/puthan-puray...
The Mathematics of India. A review by Avinash Sathaye of the book by P. P. Divakaran in the latest Bhāvanā. http://bhavana.org.in/the-mathematics-of-india/
February 9, 2026 at 12:02 AM
Prof. P. P. Divakaran (1936-2025) passed away last year. An appreciation by those who knew him.
bhavana.org.in/puthan-puray...
bhavana.org.in/puthan-puray...
Usually learn of super bowl (on some years) because I might go to some of my usual hangs, and they are filled with loud people (a bit out of character for said places). I want to say that I hate this shit, but that's too mean. In some ways, it's "better" than the spectacle around cricket matches.
February 9, 2026 at 1:39 AM
Usually learn of super bowl (on some years) because I might go to some of my usual hangs, and they are filled with loud people (a bit out of character for said places). I want to say that I hate this shit, but that's too mean. In some ways, it's "better" than the spectacle around cricket matches.
Every once in a while I will come across an arxiv submission with a particular author, and I think OK, I will get to learn about this area, even if I don't understand their result/paper. Rina Barber is one such author.
February 8, 2026 at 2:20 AM
Every once in a while I will come across an arxiv submission with a particular author, and I think OK, I will get to learn about this area, even if I don't understand their result/paper. Rina Barber is one such author.
Recently, during the course of a project, a collaborator used the Kawada-Itô theorem
www.jstage.jst.go.jp/article/ppms... Later learnt that people came up with extensions over time, and gave them names like Kawada-into www.cambridge.org/core/service... and Kawada-onto
www.jstage.jst.go.jp/article/ppms... Later learnt that people came up with extensions over time, and gave them names like Kawada-into www.cambridge.org/core/service... and Kawada-onto
www.cambridge.org
February 7, 2026 at 5:37 PM
Recently, during the course of a project, a collaborator used the Kawada-Itô theorem
www.jstage.jst.go.jp/article/ppms... Later learnt that people came up with extensions over time, and gave them names like Kawada-into www.cambridge.org/core/service... and Kawada-onto
www.jstage.jst.go.jp/article/ppms... Later learnt that people came up with extensions over time, and gave them names like Kawada-into www.cambridge.org/core/service... and Kawada-onto
Should have known about this before, but the problem formulation in so-called Bayesian optimal foraging theory is cool (with an obvious overlap with many areas in ML). An example application mitrofflab.weebly.com/uploads/2/3/...
mitrofflab.weebly.com
February 7, 2026 at 5:08 PM
Should have known about this before, but the problem formulation in so-called Bayesian optimal foraging theory is cool (with an obvious overlap with many areas in ML). An example application mitrofflab.weebly.com/uploads/2/3/...
Found today (PDF available on researchgate). Fairly interesting too, despite the title managing to promise more. dl.acm.org/doi/abs/10.5...
Toward a model of mind as a laissez-faire economy of idiots | Proceedings of the Thirteenth International Conference on International Conference on Machine Learning
dl.acm.org
February 7, 2026 at 1:47 AM
Found today (PDF available on researchgate). Fairly interesting too, despite the title managing to promise more. dl.acm.org/doi/abs/10.5...
Sometimes I really hate how long it takes me to understand stuff that should be immediate.
February 7, 2026 at 1:04 AM
Sometimes I really hate how long it takes me to understand stuff that should be immediate.
Reposted by Shubhendu Trivedi
RayRoPE: Projective Ray Positional Encoding for Multi-view Attention
Yu Wu, Minsik Jeon, Jen-Hao Rick Chang, @onceltuzel.bsky.social Shubham Tulsiani
tl;dr: even more projective geometry based PE, if you know your camera pose.
arxiv.org/abs/2601.15275
Yu Wu, Minsik Jeon, Jen-Hao Rick Chang, @onceltuzel.bsky.social Shubham Tulsiani
tl;dr: even more projective geometry based PE, if you know your camera pose.
arxiv.org/abs/2601.15275
February 6, 2026 at 1:55 PM
RayRoPE: Projective Ray Positional Encoding for Multi-view Attention
Yu Wu, Minsik Jeon, Jen-Hao Rick Chang, @onceltuzel.bsky.social Shubham Tulsiani
tl;dr: even more projective geometry based PE, if you know your camera pose.
arxiv.org/abs/2601.15275
Yu Wu, Minsik Jeon, Jen-Hao Rick Chang, @onceltuzel.bsky.social Shubham Tulsiani
tl;dr: even more projective geometry based PE, if you know your camera pose.
arxiv.org/abs/2601.15275
Reposted by Shubhendu Trivedi
If the government won't bail out my crypto losses, I'm going to become an even bigger libertarian.
February 6, 2026 at 2:04 AM
If the government won't bail out my crypto losses, I'm going to become an even bigger libertarian.
Seem like two very good papers arxiv.org/abs/2602.05862
Distribution-free two-sample testing with blurred total variation distance
Two-sample testing, where we aim to determine whether two distributions are equal or not equal based on samples from each one, is challenging if we cannot place assumptions on the properties of the tw...
arxiv.org
February 6, 2026 at 5:10 AM
Seem like two very good papers arxiv.org/abs/2602.05862
I sometimes think that badly written papers (bad upto a point) with good ideas are much more useful than well-written ones.
February 6, 2026 at 2:19 AM
I sometimes think that badly written papers (bad upto a point) with good ideas are much more useful than well-written ones.
Please ignore my typos, but the person with the simplest name here (Stefanos Pertigkiozoglou) is on the lookout for research / engineering positions. Broadly interested and very independent. Also cool as a cucumber, fun to work with/talk to.
Our recent work looking into the desgining flexible approximately equivariant NNs. arxiv.org/abs/2602.02853 This is w/ Stefanos Pertigkiozoglou (stefanospert.github.io), Mircea Petrache, Kostas Daniilidis. Builds upon our work from the 2024 edition of NeurIPS proceedings.neurips.cc/paper_files/...
Recurrent Equivariant Constraint Modulation: Learning Per-Layer Symmetry Relaxation from Data
Equivariant neural networks exploit underlying task symmetries to improve generalization, but strict equivariance constraints can induce more complex optimization dynamics that can hinder learning. Pr...
arxiv.org
February 6, 2026 at 12:25 AM
Please ignore my typos, but the person with the simplest name here (Stefanos Pertigkiozoglou) is on the lookout for research / engineering positions. Broadly interested and very independent. Also cool as a cucumber, fun to work with/talk to.
Very cool arxiv.org/abs/2602.03168
Online Conformal Prediction via Universal Portfolio Algorithms
Online conformal prediction (OCP) seeks prediction intervals that achieve long-run $1-α$ coverage for arbitrary (possibly adversarial) data streams, while remaining as informative as possible. Existin...
arxiv.org
February 5, 2026 at 11:25 PM
Very cool arxiv.org/abs/2602.03168
The irritating thing about any sort of market correction these days is that you immediately have these folk talk about how it is exactly what they predicted re: AI (link to my substack w 100k followers). Won't explain why crypto is taking a dump, or back it up by short profits --
February 5, 2026 at 8:14 PM
The irritating thing about any sort of market correction these days is that you immediately have these folk talk about how it is exactly what they predicted re: AI (link to my substack w 100k followers). Won't explain why crypto is taking a dump, or back it up by short profits --
Our recent work looking into the desgining flexible approximately equivariant NNs. arxiv.org/abs/2602.02853 This is w/ Stefanos Pertigkiozoglou (stefanospert.github.io), Mircea Petrache, Kostas Daniilidis. Builds upon our work from the 2024 edition of NeurIPS proceedings.neurips.cc/paper_files/...
Recurrent Equivariant Constraint Modulation: Learning Per-Layer Symmetry Relaxation from Data
Equivariant neural networks exploit underlying task symmetries to improve generalization, but strict equivariance constraints can induce more complex optimization dynamics that can hinder learning. Pr...
arxiv.org
February 4, 2026 at 3:50 AM
Our recent work looking into the desgining flexible approximately equivariant NNs. arxiv.org/abs/2602.02853 This is w/ Stefanos Pertigkiozoglou (stefanospert.github.io), Mircea Petrache, Kostas Daniilidis. Builds upon our work from the 2024 edition of NeurIPS proceedings.neurips.cc/paper_files/...
Reposted by Shubhendu Trivedi
Our paper will be in STOC this summer! 🌎
The simplex algorithm is super efficient. 80 years of experience says it runs in linear time. Nobody can explain _why_ it is so fast.
We invented a new algorithm analysis framework to find out.
We invented a new algorithm analysis framework to find out.
Beyond Smoothed Analysis: Analyzing the Simplex Method by the Book
Narrowing the gap between theory and practice is a longstanding goal of the algorithm analysis community. To further progress our understanding of how algorithms work in practice, we propose a new alg...
arxiv.org
February 1, 2026 at 8:35 PM
Our paper will be in STOC this summer! 🌎
By a very direct transitory relationship, more fodder for Soros conspiracy theories, if only they would notice.
February 3, 2026 at 3:13 AM
By a very direct transitory relationship, more fodder for Soros conspiracy theories, if only they would notice.
Reposted by Shubhendu Trivedi
My paper on the head direction neurons in the larval zebrafish is now published on Nature! Read it here:
www.nature.com/articles/s41...
www.nature.com/articles/s41...
Client Challenge
www.nature.com
January 10, 2026 at 11:39 AM
My paper on the head direction neurons in the larval zebrafish is now published on Nature! Read it here:
www.nature.com/articles/s41...
www.nature.com/articles/s41...
Reposted by Shubhendu Trivedi
Writing the policy gradient lecture, which gives a great opportunity to discuss the sneaky missing discount factor that makes almost every paper wrong:
arxiv.org/abs/1906.07073
arxiv.org/abs/1906.07073
Is the Policy Gradient a Gradient?
The policy gradient theorem describes the gradient of the expected discounted return with respect to an agent's policy parameters. However, most policy gradient methods drop the discount factor from t...
arxiv.org
January 29, 2026 at 3:17 AM
Writing the policy gradient lecture, which gives a great opportunity to discuss the sneaky missing discount factor that makes almost every paper wrong:
arxiv.org/abs/1906.07073
arxiv.org/abs/1906.07073
Reposted by Shubhendu Trivedi
Open postdoc position in AI in Chile (includes equivariant deep learning even if not explicitly mentioned!) cenia.cl/2026/01/27/c...
and AI + Education cenia.cl/2026/01/27/c...
and AI + Education cenia.cl/2026/01/27/c...
January 28, 2026 at 1:22 PM
Open postdoc position in AI in Chile (includes equivariant deep learning even if not explicitly mentioned!) cenia.cl/2026/01/27/c...
and AI + Education cenia.cl/2026/01/27/c...
and AI + Education cenia.cl/2026/01/27/c...