Last week I introduced this work in a domestic workshop among mathematicians and physicists, and one astrophysicist gave me a new insight to interpret the self-stabilizing potential! This is a real virtue of having a workshop with people from the communities next to us.

Han Bao@han-b.bsky.social · 2w ago

I spent the whole weekend to understand the crux of self-stabilization (arxiv.org/abs/2209.15594). To me, the most interesting part of this paper is how they model the edge of stability by ODE---because the stepsize is no longer infinitesimal at EoS and it is apparently not trivial to derive an ODE

Yeah it is really a textbook example of Goodhart's law. Giving a solution to an unsolved problem is a "proxy" toward our community's understanding of the underlying mathematical structure. Solving problems is (maybe sometimes but) not often the ultimate goal. www.economist.com/science-and-...

Top mathematicians are outraged by OpenAI’s methods

24 Fields Medal winners have written a letter of objection

economist.com

それまで人件費として支払われていたコストがAIの労働代替によって海外の寡占的なテック企業にサービスフィーとして吸い上げられ、各国の企業が雇用を通じて社会に提供していた「消費者と需要を生み出す機能」が毀損されてゆくという問題について、もっと真剣に議論されるべきだと思います。 そっちのほうがLLMはAIかどうかなんて神学論争よりずっと重要。

スドー@stdaux.bsky.social · 2w ago

「人を雇うよりAIのほうが安い」という場面も多いだろうが、問題は人を雇うと従業員の給与となって誰かの食費や学費に化けて国内消費に回るのに対して、AI使用料はどこかに消えるというのがな

I spent the whole weekend to understand the crux of self-stabilization (arxiv.org/abs/2209.15594). To me, the most interesting part of this paper is how they model the edge of stability by ODE---because the stepsize is no longer infinitesimal at EoS and it is apparently not trivial to derive an ODE

Self-Stabilization: The Implicit Bias of Gradient Descent at the Edge of Stability

Traditional analyses of gradient descent show that when the largest eigenvalue of the Hessian, also known as the sharpness $S(θ)$, is bounded by $2/η$, training is "stable" and the training loss decre...

arxiv.org

Pendant les vacances, j'ai regardé un film «Sirât» sans savoir qu'est-ce que ce film. Les paysage du désert avec des sons énormes sont magnifiques! Heureusement je pouvais le regarder à la dernière séance (il est fini au début d'août au Japon). www.imdb.com/title/tt3229...

Can't agree more. Previously one guy told me that if you want your theory paper to get accepted, the best thing is to write 30+ pages appendix (to overwhelm!). That's insane. Unless truly "necessary" (I know it's debatable), we should compress/distill formal proofs. Then we get better intuition.

People used to be able to impress and intimidate reviewers with complicated proofs. This will change. In the age of AI inscrutable proofs are cheap. It is understandable proofs that are valuable. Opaque complexity is now it is a sign of laziness or lack of insight.

GD is known to have max-margin bias, but its convergence rate is extremely slow. However, we can observe fairly decent convergence behaviors in practice as seen in the figure. I have been thinking this for a while and figured out early-stage weak convergence is possible! arxiv.org/abs/2608.04382

And the new initial meta review system is also stressful than I expected. I supposed it'd reduce rebuttal workload for both author and reviewer sides, but actually just gave us an extra burden. Very few ppl care the initial meta review and ppl do endlessly long rebuttals until the reviewers gave up🤦‍♂️

Really stressful to see AI-driven discussions during the review period this year... as an AC, I really don't have an idea how to facilitate discussion among AI-ish paper's authors and AI-ish reviewers (and the worst thing is that we cannot suppose they are truly AI even if it's very likely)

This VSCode plugin really changes my life recently! For some reason the default Claude plug-in shows LaTeX only in texts. marketplace.visualstudio.com/items?itemNa...

Claude Code LaTeX - Visual Studio Marketplace

Extension for Visual Studio Code - Adds LaTeX math rendering to Claude Code chat — inline $...$, display $$...$$, and \(...\) / \[...\]

marketplace.visualstudio.com

AI is an interesting mechanism: it leads some scientists to reveal what they would not admit willingly, that is, that they actually don't care at all about scientific rigour. AI is not the problem here, it just makes the problem visible...

way too rapid proliferation of 100% AI-generated reviews🤦‍♂️

While LLM/diffusion are extremely popular, we still can find classical topics in the poster session. This paper proposes convexified heterogeneous OT (unlike nonconvex Gromow-Wasserstein). In essence, | E[d(x,X)|y] - E[d(y,Y)|x] | is regarded as the base cost for OT. arxiv.org/abs/2606.02047

Convex Distance Operator Transport: A Convex and Geometry-Preserving Formulation

We introduce Convex Distance Operator Transport (CDOT), the first convex optimal transport framework that aligns distributions across heterogeneous domains by jointly preserving feature correspondence...

arxiv.org

I'll arrive at Seoul at Monday noon. Would love to see old friends and connect with new people there!

On Monday @ncollina.bsky.social @iraglobusharris.bsky.social and I are giving a tutorial at ICML on (multi)calibration and its applications. You can find slides and an annotated bibliography on the website: calibration-tutorial.github.io as well as an interactive demo of online calibration algs.

Calibration, Decisions, and Collaboration in Learning | ICML 2026

An ICML 2026 tutorial on making probabilistic predictions trustworthy for downstream decision-making and collaboration.

calibration-tutorial.github.io

TMLR has been facing an significant uptick in the number of submissions since the start of 2026. This is placing an extreme burden on our amazing team of reviewers & action editors. To ease this burden, TMLR will be implementing submission quotas, effective July 1. 1/n medium.com/@TmlrOrg/ann...

I participated a workshop among ML vs. PL (programming lang) people. For a long time I have zero understanding of CPS transform (en.wikipedia.org/wiki/Continu...), but a PL person told me it's essentially the dual transform from a convex set to a support function (!) Convex analysis everywhere...

It is always fantastic to read old strong papers! I read this by Koltchinskii (30 years ago) today. In this paper, he gave a convex-analytic view of the quantile function (quantile = convex conjugate supremum!), and extend it elegantly to multivariate rvs. projecteuclid.org/journals/ann...

My first PhD student Xianliang worked hard out this: In Muon, polar decomposition should always precede momentum, which significantly improves signal recovery. I'm excited to share this since few theory has been working on the benefit of momentum in Muon! arxiv.org/abs/2606.03899

Denoise First, Orthogonalize Later: Understanding Momentum in Muon via Spectral Filtering

Muon has recently demonstrated strong empirical performance in large language model training, but the theoretical role of momentum in Muon remains unclear. Existing analyses of Muon either remove mome...

arxiv.org

#AISTATS2026 It was extremely great to first see someone whose paper I have closely read, those with whom I collaborated recently without having met in-person, got invited to a next workshop, etc. Even though the time is quite challenging before the deadline😅, I really enjoyed the conference!

My new policy: if someone asks me to read something, I ask them how they used AI in creating it, and what validation/processing they applied to the AI outputs. (I disclose the same.) Been burned by giving too much attention to (undisclosed) slop folks have sent me...

D'ailleurs, j'ai pu profiter de mon séjour à Paris cette fois-ci pendant l'escale avant d'aller au Maroc. J'ai vu quelques endroits que j'aime là-bas---surtout le Centre de Pompidou, même s'il est en rénovation---aprés 7 ans! Maintenant, c'est le moment de se concentrer intensément sur le travail...

Me: I highly look forward to exploring Morocco🤩 Reality: NeurIPS😅

Half a year ago I got a quotation for a GPU server, and I redid it this month, and noticed that the server price almost doubled🤯

If you meet the eligibility requirements for the LOTUS Program and are interested in working with me, feel free to reach out. www.jst.go.jp/program/indi...

Open call for applications | LOTUS Programme

This page provides the information regarding the open call for applications for FY2025. Applications only be accepted from Japanese organizations.

jst.go.jp

Tomorrow is the very first class of my lecture at ISM (I'm gonna introduce learning theory and convex analysis). It's extremely useful for myself as well to review bunch of facts and proofs, but I need to rush because I've prepared for only half a semester😅

Sad to see more and more people rely on LLMs to generate reviews... (for some reasons, it is very easy to find at a glance; and senior people tend to rely, tho I don't intend to generalize this)

We made this public now: arxiv.org/abs/2603.10452 Hope it inspires OT enthusiasts!

Brenier Isotonic Regression

Isotonic regression (IR) is shape-constrained regression to maintain a univariate fitting curve non-decreasing, which has numerous applications including single-index models and probability calibratio...

arxiv.org

Han Bao@han-b.bsky.social · 8mo ago

Delighted to have our "Brenier isotonic regression" accepted at #AISTATS2026! We extend isotonic regression to *multiclass* setup based on Brenier optimal transport, which can be used for multiclass calibration, improving the calibration map nicely over binning etc.