Sign in

Seth Axen 🪓

@sethaxen.com
2.7K followers 372 following 230 posts

Empowering scientists with machine learning @mlcolab.org. Sometimes #Bayesian. Usually #FOSS. Often in #JuliaLang. Expat: 🇺🇸 ➡️ 🇩🇪 💼 On the job market (remote/southwest Germany) sethaxen.com

PostsRepliesMedia
Seth Axen 🪓 @sethaxen.com · 29/09/2026
My first time visiting Germany from the US, I was in awe at how punctual, quiet, and smooth the trains were (looking at you SF BART). I was surprised at how critical Germans were of DB; didn't they know how good they had it? After traveling by train in Switzerland, I get it now.
110
Reposted by Seth Axen 🪓
Aki Vehtari @avehtari.bsky.social · 09/09/2026
A FAQ is what to do in case of high Pareto-k's with PSIS-LOO or PSIS-LOGO. In case of hierarchical models we can integrate out group specific parameters before PSIS. We investigated reliability of some alternative integration methods with the eventual goal of automating this arxiv.org/abs/2609.05713
Title: Approximating Bayesian leave-one-group-out cross-validation
Authors: Anna Elisabeth Riha, Svenja Jedhoff, Paul-Christian Bürkner, Aki Vehtari
Abstract: When data are grouped, hierarchical or multilevel models are commonly used to account for group-level variation with group-specific parameters. Leave-one-group-out cross-validation (LOGO-CV) is a suitable tool for evaluating predictive performance for new groups, providing an estimator of the expected log predictive density (elpd). Brute-force LOGO-CV requires one model refit per held-out group, often using computationally expensive inference algorithms such as MCMC. This is costly, particularly for large numbers of groups or complex model structures. Commonly used importance sampling approximations, intended to reduce this cost, tend to fail because the group-specific parameters of the held-out group must be integrated out. We identify two key challenges in LOGO-CV elpd estimation: approximating the LOGO posterior and computing the grouped marginal likelihood. We compare 11 strategies, including 5 newly proposed, to address them. Among others, we combine Pareto-smoothed importance sampling or adaptive importance sampling with integration techniques such as Laplace approximation, adaptive Gauss-Hermite quadrature, and bridge sampling. We evaluate these strategies in both simulation experiments and real-world case studies, which show that marginalising over the group-specific parameters substantially improves the reliability of the importance sampling approaches.
1233
Seth Axen 🪓 @sethaxen.com · 21/07/2026
Great write-up and comparison! I'd also completely forgetting about my original idea and write-up, so thanks @spinkney.bsky.social for improving it and showing it works well in practice!
270
Seth Axen 🪓 @sethaxen.com · 21/07/2026
I don't post here very often, so I think it's worth checking in here to note that since I wrote this, I've transitioned to a heavy agent-based dev workflow. Partially that's due to a huge improvement in the models, partially due to some useful skills/harnesses.
1341
Seth Axen 🪓 @sethaxen.com · 17/07/2026
These PhD projects look like a ton of fun!
000
Seth Axen 🪓 @sethaxen.com · 08/07/2026
Bring 👏back 🦶blue books! 📘
010
Seth Axen 🪓 @sethaxen.com · 26/06/2026
That feeling when you join a weekly video group check-in with your strength training coach, and the very first word you hear is "Bayesian"... 🤗
120
Reposted by Seth Axen 🪓
TMLR Published Papers @tmlr-pub.bsky.social · 09/04/2026
Amortized Bayesian Workflow Chengkun LI, Aki Vehtari, Paul-Christian Bürkner et al. Action editor: Tom Rainforth openreview.net/forum?id=osV7adJlKD #mcmc #generative #amortized
033
Reposted by Seth Axen 🪓
ArviZ @arviz.bsky.social · 19/03/2026
ArviZ 1.0 is out! We have refactored it to be more modular, flexible & lightweight. For an overview of the changes, check the migration guide. python.arviz.org/en/stable/us...
python.arviz.org
Getting started — ArviZ 1.0.0 documentation
21510
Reposted by Seth Axen 🪓
Rafe Meager (they/them) @economeager.bsky.social · 12/03/2026
"Other extensions of the factorial function do exist, but the gamma function is the most popular and useful. " did the gamma function herself/itself write this
2638
Seth Axen 🪓 @sethaxen.com · 12/03/2026
Anyone aware of a collection of agent skills tailored towards FOSS workflows for package development and maintenance? Strongly considering writing my own.
111
Reposted by Seth Axen 🪓
Aki Vehtari @avehtari.bsky.social · 04/03/2026
If you have been using LOO-PIT, this is a must read for you! @herman-tesso.bsky.social has done excellent work with this paper! Thanks for @florencebockting.bsky.social and @aloctavodia.bsky.social for getting this to bayesplot and ArviZ. I'll notify when I have my casestudies updated with this
0162
Seth Axen 🪓 @sethaxen.com · 28/01/2026
I've been seeing some folks recently explain how they're using agent-based workflows to help manage their life/work. I think it's great that people are sharing this and probably super useful for some. But I still haven't seen a use case that would be helpful for me.
141
Seth Axen 🪓 @sethaxen.com · 07/01/2026
Over the break I was picking away at a personal project where brms is clearly the best tool to use, and hot damn, is brms nice!
0101
Seth Axen 🪓 @sethaxen.com · 03/01/2026
Introduce yourself as what almost killed you: Hi there, I'm a hernia
010
Reposted by Seth Axen 🪓
Demetri @phdemetri.bsky.social · 26/11/2025
I believe eigen fly. I believe eigen touch the sky.
324042
Seth Axen 🪓 @sethaxen.com · 24/11/2025
I was pleasantly surprised to find that a TMLR paper last year (openreview.net/forum?id=Kre...) cited in one of their proofs my blog post on injectivity radii for unitary groups (sethaxen.com/blog/2023/02...).
openreview.net
Wrapped $\beta$-Gaussians with compact support for exact...
We introduce wrapped $\beta$-Gaussians, a family of wrapped distributions on Riemannian manifolds, supporting efficient reparametrized sampling, as well as exact density estimation, effortlessly...
120
Seth Axen 🪓 @sethaxen.com · 24/11/2025
Finally got around to trying out @typst.app, and I'm really surprised how easy the learning curve coming from TeX has been! I'm still not convinced it has all of the features I would want to replace TeX for papers, but it might replace my current TeX-in-MD derivation workflow.
290
Seth Axen 🪓 @sethaxen.com · 21/11/2025
Updating CUDA on my desktop machine today, thoughts and prayers appreciated.
1130
Seth Axen 🪓 @sethaxen.com · 19/11/2025
I've been experimenting with using Conventional Commits on some of my repos, and I think I'll start using it everywhere. I found it helps me structure my self-contained commits and spot when which changes were made where more easily. www.conventionalcommits.org/en/v1.0.0/
conventionalcommits.org
Conventional Commits
A specification for adding human and machine readable meaning to commit messages
280
Reposted by Seth Axen 🪓
ML for Science @ml4science.bsky.social · 13/11/2025
Great new work from the labs of @jakhmack.bsky.social and @philipp.hertie.ai! The software Jaxley enables brain simulations which both imitate the processes in the brain in detail and can solve challenging cognitive tasks. Press release of @unituebingen.bsky.social: uni-tuebingen.de/en/universit...
uni-tuebingen.de
Software optimizes simulations of the brain
0164
Reposted by Seth Axen 🪓
BenjMurrell @benjmurrell.bsky.social · 10/11/2025
We figured out flow matching over states that change dimension. With "Branching Flows", the model decides how big things must be! This works wherever flow matching works, with discrete, continuous, and manifold states. We think this will unlock some genuinely new capabilities.
42413
Reposted by Seth Axen 🪓
Diana Cai @dianarycai.bsky.social · 27/10/2025
Fisher meets Feynman! 🤝 We use score matching and a trick from quantum field theory to make a product-of-experts family both expressive and efficient for variational inference. To appear as a spotlight @ NeurIPS 2025. #NeurIPS2025 (link below)
Fisher meets Feynman: score-based variational inference with a product of experts
1469
Seth Axen 🪓 @sethaxen.com · 24/10/2025
Yesterday I asked the 7-year-old how he would describe what I do for work. His response: "You do math all day, get paid for no reason, and print things."
160
Seth Axen 🪓 @sethaxen.com · 02/10/2025
"I'm sure I don't have to tell you that #JuliaLang is awesome and Matlab is not" #HeardAtJuliaCon #JuliaCon
2211
Reposted by Seth Axen 🪓
MC Stan @mc-stan.org · 17/09/2025
MC Stan is here! Follow for the latest Stan news, and tag if you want us to repost your posts about new papers, packages, courses, etc. about Stan
13831
Reposted by Seth Axen 🪓
sbi - Simulation-based inference @sbi-devs.bsky.social · 09/09/2025
From hackathon to release: sbi v0.25 is here! 🎉 What happens when dozens of SBI researchers and practitioners collaborate for a week? New inference methods, new documentation, lots of new embedding networks, a bridge to pyro and a bridge between flow matching and score-based methods 🤯 1/7 🧵
12916
Reposted by Seth Axen 🪓
Charles Margossian @charlesm993.bsky.social · 10/09/2025
My paper with Loucas Pillaud-Vivien and Lawrence Saul, “Variational Inference for Uncertainty Quantification: An Analysis of Trade-offs”, has been accepted for publication in the Journal of Machine Learning Research. 📃 arxiv.org/abs/2403.13748 🧵 1/
1165
Seth Axen 🪓 @sethaxen.com · 08/09/2025
Working from home and overhearing my 7-year-old trying to explain "countable infinity" to my partner using donuts.
240
Reposted by Seth Axen 🪓
Jan Boelts @janboelts.bsky.social · 18/08/2025
Fun read of their amazing contributions to the SBI hackathon! 🥐 The SBI-Pyro bridge that @sethaxen.com built has a lot of potential I believe. I'll actually be presenting this work at @euroscipy.bsky.social this Wednesday - excited to share this with a broader audience. euroscipy.org/talks/KCYYTF/
euroscipy.org
Pyro Meets SBI: Unlocking Hierarchical Bayesian Inference for Complex Simulators
The EuroSciPy meeting is a cross-disciplinary gathering focused on the use and development of the Python language in scientific research.
051
Reposted by Seth Axen 🪓
Aki Vehtari @avehtari.bsky.social · 02/08/2024
All three books I've co-authored are freely available online for non-commercial use: - #Bayesian Data Analysis, 3rd ed (aka BDA3) at stat.columbia.edu/~gelman/book/ - #Regression and Other Stories at avehtari.github.io/ROS-Examples/ - Active Statistics at avehtari.github.io/ActiveStatis...
The cover of Bayesian Data Analysis bookThe cover of Regression and Other Stories bookThe cover of Active Statistics book
7368145
Seth Axen 🪓 @sethaxen.com · 11/08/2025
Pro-tip: don't be the poor sod that adds a daily cron trigger to a GitHub Actions workflow that often fails. Long after you've left the project, you will get *daily* failure notifications, and the only way out is to trick some other poor sod into editing the workflow. docs.github.com/en/actions/c...
010
Seth Axen 🪓 @sethaxen.com · 08/08/2025
Like the authors, I also found this result disturbing. The crux is that *conditioning* a distribution to lie on a manifold is *not* in general the same thing as *restricting* the distribution to the manifold (i.e. constraining the support and re-normalizing).
140
Seth Axen 🪓 @sethaxen.com · 08/08/2025
For the Bayesians, it's the Monte Hall problem.
030
Seth Axen 🪓 @sethaxen.com · 05/08/2025
Finally, a part of my Fairphone 5 stopped working. For any other phone, I'd be dropping >500€ right now to replace it with a marginally better brand new phone that would last me just 2-3 years. Instead, I paid 41€, had the new part in 4 days, replaced it in 5 minutes, and the phone is good as new!
150
Seth Axen 🪓 @sethaxen.com · 31/07/2025
Sharing this here a bit late, but @vstaros.bsky.social and I wrote a little something about our experience contributing to the @sbi-devs.bsky.social (simulation-based inference) hackathon. @mlcolab.org @mackelab.bsky.social We were obviously very hungry while writing.
mlcolab.org
A retrospective on the 2025 SBI Hackathon
You walk into a bakery, take one bite of a still-warm pastry, and think: “Whoa - there’s rye flour, a hint of orange zest, maybe cardamom… and is that buckwheat honey?” From that single taste you begi...
0113
Seth Axen 🪓 @sethaxen.com · 24/07/2025
Spent a pleasant morning organizing GitHub repos and notifications, responding to issues and PRs, and answering questions on Slack. Reminder that FOSS is often about building a community as much as building software!
170
Reposted by Seth Axen 🪓
Martin Trapp @trappmartin.eurosky.social · 21/07/2025
Remember that computers use bitstrings to represent numbers? We exploit this in our recent @auai.org paper and introduce #BitVI. #BitVI directly learns an approximation in the space of bitstring representations, thus, capturing complex distributions under varying numerical precision regimes.
BitVI on 1D Gaussian mixture models.
2223
Seth Axen 🪓 @sethaxen.com · 15/07/2025
If you run a DuckDuckGo search for "qr decomposition", it returns a QR code for "decomposition".
1110
Reposted by Seth Axen 🪓
Michael Kirchhof @mkirchhof.bsky.social · 03/07/2025
Can LLMs access and describe their own internal distributions? With my colleagues at Apple, I invite you to take a leap forward and make LLM uncertainty quantification what it can be. 📄 arxiv.org/abs/2505.20295 💻 github.com/apple/ml-sel... 🧵1/9
1236
Seth Axen 🪓 @sethaxen.com · 20/06/2025
This paper is an absolute work of art.
121
Seth Axen 🪓 @sethaxen.com · 07/06/2025
Normalize v1.0 of software adding no new features. 1.0 comes with an expectation of greater stability than pre-1.0 releases. Debuting new features in that release makes it less stable, not more. Put the new features in an earlier release and let them stabilize *before* the 1.0.
0243
Seth Axen 🪓 @sethaxen.com · 02/06/2025
I strongly suspect my 4-year-old has figured out how to vomit at will so he can stay home from preschool and "watch TV."
020
Reposted by Seth Axen 🪓
Philipp Berens @philipp.hertie.ai · 23/05/2025
We are incredible happy to be able to continue our work of developing new #AI4science across a wide range of disciplines with incredible colleagues in #physics, #neuroscience, #cogsci, #geoscience, #linguistics, #economics, #medicine, #philosophy, #law and #anthropology! @unituebingen.bsky.social
1395
Seth Axen 🪓 @sethaxen.com · 23/05/2025
I am literally begging you, if in your paper you summarize the features of software *that you do not use*, at the bare minimum have a user of that software (even better, a maintainer) look over that bit of text before submitting. Sincerely, tired-of-software-I-worked-on-being-misrepresented
240
Reposted by Seth Axen 🪓
Claire Vernade @claireve.bsky.social · 04/05/2025
Amazing Best Paper award talk on Variational inference theory by Charles Margossian @charlesm993.bsky.social at #aistats What does VI learn and under which conditions? -> arxiv.org/pdf/2410.11067
0162
Seth Axen 🪓 @sethaxen.com · 28/04/2025
From Student's "Errors of Routine Analysis" (1927) doi.org/10.2307/2332...
032
Reposted by Seth Axen 🪓
Aki Vehtari @avehtari.bsky.social · 23/04/2025
A new paper with Alex Cooper and Catherine Forbes "Joint leave-group-out cross-validation in Bayesian spatial models" arxiv.org/abs/2504.15586 (Alex did the hard work for this, and running many cross-validation simulations with spatial models is hard)
Abstract: Cross-validation (CV) is a widely-used method of predictive assessment based on repeated model fits to different subsets of the available data. CV is applicable in a wide range of statistical settings. However, in cases where data are not exchangeable, the design of CV schemes should account for suspected correlation structures within the data. CV scheme designs include the selection of left-out blocks and the choice of scoring function for evaluating predictive performance. This paper focuses on the impact of two scoring strategies for block-wise CV applied to spatial models with Gaussian covariance structures. We investigate, through several experiments, whether evaluating the predictive performance of blocks of left-out observations jointly, rather than aggregating individual (pointwise) predictions, improves model selection performance. Extending recent findings for data with serial correlation (such as time-series data), our experiments suggest that joint scoring reduces the variability of CV estimates, leading to more reliable model selection, particularly when spatial dependence is strong and model differences are subtle.
1195
Seth Axen 🪓 @sethaxen.com · 14/04/2025
Phew! US Taxes filed. While they're generally not fun to do, it's extra not fun when you're an expat and only have to do it to avoid being double-taxed while all the immigrants from other countries don't have to file taxes in their home countries.
020
Seth Axen 🪓 @sethaxen.com · 04/04/2025
o1 musing to itself in German: "Sister, I am working on considering all possibilities."
2113