Rémi Flamary

These results are crazy and so is any reviewer that do not find them impressive. Also making AI more data efficient and actually accessible for non major companies and non state actors is important for us all. Keep fighting good science always wins in the end.

Hum I have some scores disappearing on openreview on my #NeurIPS papers. Do you have the same? Maybe it means the scaore has been edited? If this is to hide the score until final decision, this seems like a bad idea (we know how to do data imputation in the community)

Everybody misses the 1-page rebuttal.

These lengthy forum style comments are a nightmare: a nightmare for the authors who spend way too much time writing them, a nightmare for the reviewers who spend too much time understanding them, a nightmare for the ACs who will have to summarize all. Stop it!

As a #NeurIPS AC I did a little stat. Paper reviews+responses in my batch contain 515k characters and 75k words. This is without including papers themselves or authors/reviewers discussions. So please be nice to your AC, we are doing our best. Also I miss the one page PDF response of old ICML.

We should do research that we think is important and not that the reviewers will like. It I'll be accepted in the end but I agree that it's becoming harder with noisy reviews and publication expectations for PhD students are ridiculously high nowadays so it's a hard choice.

This prompt injection for GEMINI CLI is just fascinating. It's crazy that because of LLM, hackers can now do social engineering without a human in the loop. But I guess giving execution rights to an LLM is already a big humain failure anyways. arstechnica.com/security/202...

There is a huge bias between established researchers and reknown teams and others in terms of traction on arxiv. This would kill diversity and only promote trending topics I think.

Finally most important, thanks to my awesome collaborators : Yanis Lalou, @tgnassou.bsky.social @antoinecollas.bsky.social , Antoine de Mathelin, @ambroiseodt.bsky.social , Thomas Moreau, Alexandre Gramfort and all the SKADA contributors 14/14 scikit-adaptation.github.io

I'm very proud of this work that shows that despite impressive performances of DA in the literature we still need to solve import questions such as practical validation before people can trust it in production pipelines 13/n

But the best thing is that since it is 100% reproducible and we provide the results files on our github you do no need to re-run it on all methods on your servers if you just want to compare your new DA method. You can just run your new method and use pre-computed performances on competitors. 12/n

SKADA-bench is implemented using SKADA DA toolbox and Benchopt to ensure distributed run on slurm servers and 100% reproducibility. You can add a new DA methods or datasets with just a few lines of code. 11/n github.com/scikit-adapt...

We also provide computational time for all methods (including with validation time) that shows that even fast methods can be very long to validate when they have multiple hyperparameters. 10/n

We also provide in supplementary as study of the loss of performance due to realistic validation (DA Scorer) VS (unrealistic DA using target labels) and find that some methods can loose from 5 to 10% accuracy in this case. 9/n

We also compared several Deep DA methods based on invariant features representations and show that they are as expected better than shallow on Vision DA but can fail on other modalities such as biomedical signals (and be worst than shallow methods). 8/n

Circular Validation consists in performing DA from Source to Target and use the estimated target classifier to label target samples, then performing DA from target to source, and compute accuracy on source (label preservation). It was proposed in DASVM paper 7/n rslab.disi.unitn.it/papers/R82-P...

We also provide a comparison of different realistic DA scorers and plot their correlation wrt true target accuracy. Interestingly the best one across models and modalities is one of the oldest but rarely used in practice : Circular validation 6/n

One nice result is that linear alignment methods (CORAL and Linear Optimal Transport) that have few parameters seem to work well on real data and are the only methods that do not decrease performance wrt no DA across modalities. 5/n

Results on shallow DA show that methods designed for specific shifts work as expected on them on simulated data. But on real data adaptation performance (and if better than training on source data) is very dataset and method dependent. Many DA methods perform worst than no DA. 4/n

The benchmark uses a nested cross-validation with realistic DA scorers that do not use target data labels (since they are unavailable in practice) to select the methods hyperparameters. This is a reality check for more complex DA approaches that are difficult to validate in practice. 3/n

We compare 21 Shallow DA methods and 7 Deep DA methods on multiple modalities including simulated shifts, images, text, tabular and biomedical signals. For shallow methods we used pre-trained deep embeddings to evaluate their ability to adapt in the feature space. 2/n

SKADA-Bench : Benchmarking Unsupervised Domain Adaptation Methods with Realistic Validation On Diverse Modalities, has been published published in TMLR today 🚀. It was a huge team effort to design (and publish) an open source fully reproducible DA benchmark 🧵1/n. openreview.net/forum?id=k9F...

I love the #eurips initiative! But to live up to its potential, it should be accepted an official #neurips conference location (similar to Mexico City) and not just an addon! Then we would save C02, rather than adding to it!!! What can we do to achieve this?

My academic job offer was rescinded. I’ll keep going—but U.S. researchers are running out of road Despite an uncertain future, this postdoc will “keep doing the science I love while I still have a bench”

This is awful! We need to treat our scientists better if we want to do good science. This means consistent funding for tenure that allows time for risky but novel discoveries and not only short-time projects and infinite post-docs. www.science.org/content/arti...

I seem to remember a dinosaur movie where the annoying skeptic keeps telling people it was not a good idea to make new ones.

✨Thrilled to see EurIPS launch — the first officially endorsed European NeurIPS presentation venue!

👀 But NeurIPS now requires at least one author to attend in San Diego or Mexico (and not just virtually as before). This is detrimental to many. Why not allow presenting at EurIPS or online?
1/4

A meme featuring Gru who wonders about NeurIPS’s plan to allow a satellite event in Europe but also force presentation in the US

Actually it is unclear to me from the post and website if you can do only one and present at eurips instead of Neurips and how it works in terms of registration. Does anyone have more information?

It is a very nice step toward distributed conference that we really need. Big thanks to organizers. But I agree that having to go to both is a hard constraint and also will be a net negative from an environmental perspective.

Edwige Cyffer presenting her work on the interaction between decentralization and privacy after receiving the 2025 best PhD prize from the French machine learning scientific society (SFFAM) at CAP 2025.

@rflamary.bsky.social 's keynote at CAp'25 was a blast!

A short summary in the linked post below...