Daniel Csillag
LightNews LightNews
banner
dccsillag.xyz
Daniel Csillag
@dccsillag.xyz
850 followers 380 following 62 posts
Applied mathematician working on machine learning, statistics and compilers. Currently doing research at FGV EMAp.

dccsillag.xyz
Posts Replies Media Videos
dccsillag.xyz
Off to ICML soon! Excited to talk all things ML&stats: UQ, e-values, safety, optimization, etc
If you'd like to meet, shoot me an email!
July 11, 2025 at 2:19 AM
dccsillag.xyz
It's good to finally have a good reference for this stuff! Kudos to the authors.
arxiv.org/abs/2501.18374
Proofs for Folklore Theorems on the Radon-Nikodym Derivative
In this paper, rigorous statements and formal proofs are presented for both foundational and advanced folklore theorems on the Radon-Nikodym derivative. The cases of conditional and marginal probabili...
arxiv.org
April 25, 2025 at 3:04 PM
dccsillag.xyz
I'll be presenting Strategic Conformal Prediction at AISTATS next week.

Looking forward to chatting about all things uncertainty quantification!
dccsillag.xyz Daniel Csillag @dccsillag.xyz · Feb 4
Happy to announce that our paper, 𝗦𝘁𝗿𝗮𝘁𝗲𝗴𝗶𝗰 𝗖𝗼𝗻𝗳𝗼𝗿𝗺𝗮𝗹 𝗣𝗿𝗲𝗱𝗶𝗰𝘁𝗶𝗼𝗻, is now accepted to AISTATS 2025!

Is your uncertainty quantification robust to people trying to break it?
Ours is :)
Front page of our paper, 'Strategic Conformal Prediction'. Abstract: When a machine learning model is deployed, its predictions can alter its environment, as better informed agents strategize to suit their own interests. With such alterations in mind, existing approaches to uncertainty quantification break. In this work we propose a new framework, Strategic Conformal Prediction, which is capable of robust uncertainty quantification in such a setting. Strategic Conformal Prediction is backed by a series of theoretical guarantees spanning marginal coverage, training-conditional coverage, tightness and robustness to misspecification that hold in a distribution-free manner. Experimental analysis further validates our method, showing its remarkable effectiveness in face of arbitrary strategic alterations, whereas other methods break.
April 24, 2025 at 3:33 PM
dccsillag.xyz
Happy to announce that our paper, 𝗦𝘁𝗿𝗮𝘁𝗲𝗴𝗶𝗰 𝗖𝗼𝗻𝗳𝗼𝗿𝗺𝗮𝗹 𝗣𝗿𝗲𝗱𝗶𝗰𝘁𝗶𝗼𝗻, is now accepted to AISTATS 2025!

Is your uncertainty quantification robust to people trying to break it?
Ours is :)
Front page of our paper, 'Strategic Conformal Prediction'. Abstract: When a machine learning model is deployed, its predictions can alter its environment, as better informed agents strategize to suit their own interests. With such alterations in mind, existing approaches to uncertainty quantification break. In this work we propose a new framework, Strategic Conformal Prediction, which is capable of robust uncertainty quantification in such a setting. Strategic Conformal Prediction is backed by a series of theoretical guarantees spanning marginal coverage, training-conditional coverage, tightness and robustness to misspecification that hold in a distribution-free manner. Experimental analysis further validates our method, showing its remarkable effectiveness in face of arbitrary strategic alterations, whereas other methods break.
February 4, 2025 at 12:59 PM
Reposted by Daniel Csillag
andrewgwils.bsky.social
We're excited to announce the ICML 2025 call for workshops! The CFP and submission advice can be found at: icml.cc/Conferences/.... The deadline is Feb 10. Submit some creative proposals!
ICML 2025 Call for Workshops
icml.cc
January 3, 2025 at 2:19 AM
Reposted by Daniel Csillag
auai.org
The 41st Conference on #Uncertainty in #AI will be held in Rio de Janeiro 🇧🇷, July 21-25!

The CfP is out 👉 www.auai.org/uai2025/call...

🚨 Feb 10: Paper submission
🗣️ Apr 3-10: rebuttal period
🎉/💀 May 6: Author notification

#UAI2025 #ML #stats #learning #reasoning #uncertainty
December 3, 2024 at 5:07 PM
Reposted by Daniel Csillag
scholzmx.bsky.social
Ezpz #rstats
Rick and Morty standing in front of a portal. The "20 minutes, what's the worst that could happen" moment. Text says "me, about to tune some hyper parameters"
December 21, 2024 at 7:39 PM
Reposted by Daniel Csillag
omarrivasplata.bsky.social
📢 Outstanding PhD student wanted! 🤓

The successful candidate will be based at The University of Manchester Dept. of CS ✨ to work on learning theory and methods for novel types of distributional shifts, co-supervised with @samikaski.bsky.social

⏳ DL 31.Jan.2025.
www.findaphd.com/phds/project...
Learning theory and methods for novel types of distributional shifts. at The University of Manchester on FindAPhD.com
PhD Project - Learning theory and methods for novel types of distributional shifts. at The University of Manchester, listed on FindAPhD.com
www.findaphd.com
December 15, 2024 at 5:45 PM
dccsillag.xyz
Everyone should know how to do binary search and bracketing
December 13, 2024 at 11:29 AM
dccsillag.xyz
Off to NeurIPS soon!
December 8, 2024 at 10:40 AM
Reposted by Daniel Csillag
huytransformer.com
go.bsky.app/7U9DDTv
November 20, 2024 at 12:07 PM
dccsillag.xyz
@johndcook.bsky.social I just read your recent 'Categorical Data Analysis' post www.johndcook.com/blog/2018/04....
As a ML&stats person who tried to incorporate category theory in my work, I'd say the problem is that it just doesn't look like CT adds much. Whenever I try to incorporate it, ..
Categorical Data Analysis
Categorical data analysis could mean a couple different things. One is analyzing data that falls into unordered categories (e.g. red, green, and blue) rather than numerical values (e.g. height in cent...
www.johndcook.com
November 23, 2024 at 10:58 AM
Reposted by Daniel Csillag
yongyuanxi.bsky.social
Many think Rust is primarily about memory safety, but the real reason to use Rust is developer productivity (for C++-like system programming) from having a sane package manager and type system.
November 21, 2024 at 12:30 PM
Reposted by Daniel Csillag
trappmartin.bsky.social
For those who don’t know yet, I am organising an online talk series together with Arno Solin on “Advances in Probabilistic Machine Learning (APML)”.

It’s free for everyone to join and support early career researchers!

You can register and check out the schedule here: aaltoml.github.io/apml/
Seminar on Advances in Probabilistic Machine Learning
This seminar series aims to provide a platform for young researchers (PhD student or post-doc level) to give invited talks about their research, intending to have a diverse set of talks & speakers on ...
aaltoml.github.io
November 20, 2024 at 8:33 PM
dccsillag.xyz
Anthropic published a paper telling people to use confidence intervals for evals.

I now await for their next paper, which will explain multiple comparisons to the LLM people
November 21, 2024 at 6:20 PM
Reposted by Daniel Csillag
arxiv-stat-ml.bsky.social
Daniel Csillag, Claudio Jos\'e Struchiner, Guilherme Tegoni Goedert
Strategic Conformal Prediction
https://arxiv.org/abs/2411.01596
November 5, 2024 at 5:30 AM
dccsillag.xyz
NeurIPS registrations following a *randomized lottery* is absurd to me. Especially once you consider all the logistics involved with attendance (e.g., flights, stay, etc.)
October 27, 2024 at 4:11 AM
Reposted by Daniel Csillag
rafael-izbicki.bsky.social
Ran my text through ChatGPT arguing why oversampling usually doesn’t make sense, and it flipped my argument to support oversampling instead. A reminder that AI often defaults to consensus—even when it's wrong. Always double-check!

#AI #MachineLearning #Statistics
September 21, 2024 at 11:38 AM
dccsillag.xyz
tqdm really feels like the best library ever written in the history of ever. It's so good...
September 10, 2024 at 3:41 PM
Reposted by Daniel Csillag
ohmygoedel.bsky.social
A thread about some of my favorite books:

0. Logic and Structure. Dirk van Dalen.
Logic and Structure. Dirk van Dalen.
June 14, 2023 at 2:50 PM
dccsillag.xyz
Gradient descent with random initialization avoids troublesome stationary points almost surely.

#🧮
Let $f : RR^n -> RR$ be twice continuously differentiable and assume its gradient is Lipschitz. Consider gradient descent with random initialization: $ x_0 ~ nn(bold(0),I), wide wide x_(t+1) = x_t - eta nabla f(x_t). $ If the procedure converges to some point $macron(x)$ and $eta$ is sufficiently small, then, almost surely: (i) It is a critical/stationary point: $nabla f(macron(x)) = 0$. (ii) It is a second-order stationary point: the Hessian $nabla^2 f(macron(x))$ is positive-semidefinite. This means that it will not converge to maxima, nor most other spurious stationary points. So, even though GD only uses first-order information (the gradient), we get a second-order guarantee "for free!"
September 4, 2024 at 4:29 AM
dccsillag.xyz
Useful fact: you can bound the gap between any two probability distributions using f-divergences. A rather nice case is with the Pearson chi-squared f-divergence.
September 1, 2024 at 11:57 PM
dccsillag.xyz
Just arrived here! Liking bsky so far, we'll see how it goes.

Follow for occasional posts on math, ML and stats :)
September 1, 2024 at 9:45 PM