Quentin Berthet

Being a rejected paper seems like a good criteria (possible issues, but very easy to implement)

I'm surprised, there must have been an additional round of changes.

All the emails I saw as an SAC were "help your ACs write new meta-reviews".

I think that they are hard to bootstrap.

While I agree that general chairs, program chairs, local chairs etc do a lot of work (seeing a glimpse of that sometimes myself with ALT/AISTATS), once you start having a bit of money, it can get easier with using conference organising services

How do-able would it be to organize a conference with its deadline a week after NeurIPS, held a week before, that would always take place physically in Europe (but welcome contributions from anywhere)?

I'm very happy that EurIPS is happening, and planning to attend.

But as a NeurIPS SAC I must say that the last definition should not exist - ACs are supposed to change the decisions and update their meta-review themselves.

🚨AISTATS workshop proposals open!

Big News: For the first time, there will be a day of workshops at #AISTATS 2026, in Tangier, Morocco 🌴🇲🇦

Quentin Berthet @qberthet.bsky.social and I are workshop chairs.

virtual.aistats.org/Conferences/...
Deadline: Oct 17, AOE

At #AISTATS2025, I will be giving an "Oral" presentation of our work on "Implicit Diffusion"

arxiv.org/abs/2402.05468

I will be attending #ICLR2025 in Singapore and #AISTATS2025 in Mai Khao over the next two weeks.

Looking forward to meeting new people and learning about new things. Feel free to reach out if you want to talk about Google DeepMind.

ICML reviews are out

Very proud to have contributed to this work, and very happy to have learnt a lot about votes, choices, and agents!

Il y a plus de 200 notes ?

🇨🇦

📣 New preprint 📣

Learning Theory for Kernel Bilevel Optimization

w/ @fareselkhoury.bsky.social E. Pauwels @michael-arbel.bsky.social

We provide generalization error bounds for bilevel optimization problems where the inner objective is minimized over a RKHS.

arxiv.org/abs/2502.08457

These findings are quite consistent

Our end-to-end method captures a regression and a classification objective, as well as the autoencoder loss.

We see it as "building a bridge" between these different problems.

8/8

This is a simple generalization from previous binning approaches, the main difference being that we learn the encoding.

We compare different training methods, showing up to 25% improvement on the least-squares baseline error for our full end-to-end method, over 8 datasets.

7/8

You can then train a classification model π_θ (green, top) on the encoded targets, with a KL/cross-entropy objective!

At inference time, you use the same decoder μ to perform your prediction!

6/8

First, the architecture - we use a general target encoder Ψ_w (red, bottom) that transforms the target y in a distribution over k classes

e.g. use softmax(dist) to k different centers

The encoder and the associated decoder μ (in blue) can be trained on an autoencoder loss

5/8

In this work, we propose improvements for "regression as classification"

- Soft-binning: encode the target as a probability, not just a one-hot.
- Learnt target encoders: Instead of designing this transformation by hand, learn it from data.
- Train everything jointly!

4/8

The idea is to transform each value of the target into a class one-hot, and to train a classification model to predict the value of the target. (here with y in 2D, with a grid)

It seems strange, but it's been shown to work well in many settings, even for RL applications.

3/8

An intriguing observation:

In many tasks with a continuous target (price, rating, pitch..), instead of training on a regression objective with least-squares [which seems super natural!] - people have been instead training their models using classification!

2/8

Building Bridges between Regression, Clustering, and Classification Regression, the task of predicting a continuous scalar target y based on some features x is one of the most fundamental tasks in machine learning and statistics. It has been observed and...

Check out our paper, with Lawrence Stewart and @bachfrancis.bsky.social

Link: arxiv.org/abs/2502.02996

1/8

🚨 New paper on regression and classification!

Adding to the discussion on using least-squares or cross-entropy, regression or classification formulations of supervised problems!

A thread on how to bridge these problems:

Brian Driscoll, the acting director of the FBI, sporting curly hair, a kinda rakish jazz dirtbag moustache and beard combo, a faintly amused look in his eyes and a wry "whaddaya gonna do" smirk

I've got to say, the guy who accidentally became the Director of the FBI does 100% look like the guy who accidentally becomes the Director of the FBI in a mid-2000s comedy about a guy who accidentally becomes the Director of the FBI

🚀 Policy gradient methods like DeepSeek’s GRPO are great for finetuning LLMs via RLHF.

But what happens when we swap autoregressive generation for discrete diffusion, a rising architecture promising faster & more controllable LLMs?

Introducing SEPO !

📑 arxiv.org/pdf/2502.01384

🧵👇

Congratulations!

The only "nice" thing is that if you have a lot of centered equations that are not very wide, it saves space. This is very visible when you switch templates e.g. from ICML to NeurIPS.

For everything else, it's just horrible. No surprise that "unsubmitted preprints" in ML are single column

Theorem. If $n$ is an integer and $n^2$ is even, then $n$ is itself even. Proof. Contrapositives are for cowards, so assume $n$ is an integer and $n^2$ is even. Then $n^2=2k$ for some integer $k$, and thus $n^2-2k=0$. Behold: \[ n = n + (n^2-2k) = n(n+1)-2k. \] Both $n(n+1)$ and $2k$ are even, so $n$ is even. QED.

Contrapositives are for cowards. Behold.

#AISTATS results are out

Also, researchers who don't have a Fields medal might feel more pressure to travel to advertise their work and network with others.