Come join us in a great department in a great city! I'm not on the search committee, but feel free to be in touch if you have any questions or anything.

REMINDER!! We hope to see you all ✨TODAY✨ for our social! Check out queerinai.com/icml-2025 for further details! 🌈🌈

1/ 💻 Queer in AI is hosting a social at #ICML2025 in Vancouver on 📅 July 16, and you’re invited! Let’s network, enjoy food and drinks, and celebrate our community. Details below…

As things get ever worse in the UK and the US releases sloppy political “reviews” of bad science (magnified by the NYT/Atlantic/etc), a good reminder for anyone who has “concerns” about trans youth care.

(Canada seems okay federally for now, but AB/SK are going in v bad directions.)

Screenshot of a slide: AI Scientists still suffer from hallucinations “We use an LSTM-based neural network (Goodfellow et al. 2016)” is highlighted, with the comment “Should be Hochreiter and Scmidhuber”

AI Scientist making the funniest possible citation mistake (Goodfellow instead of Schmidhuber)

Queer in AI @ ICLR 2025: Social Sign-In Survey Please note that most fields below are optional. We also strongly encourage you to fill out the Queer in AI Demographics Survey which provides QinAI with information about our community we use to help...

Join us tomorrow for our official social at #ICLR2025 on Saturday, 04/26, 5:00 - 6:30 pm in the GHJ room ("hidden" between Hall 2 and 3)! We can’t wait to see you all and to mingle! After the social ends, we will also set out for an unofficial bar tour of downtown Singapore.

Thrilled to announce that Joshua’s paper won one of three Outstanding Paper awards at ICLR.

Come to the poster on Friday afternoon (#376 in Hall 3) or the talk on Saturday (4:30 in Hall 1), and while you’re at it snag him for a postdoc!

📢Curious why your LLM behaves strangely after long SFT or DPO?
We offer a fresh perspective—consider doing a "force analysis" on your model’s behavior.
Check out our #ICLR2025 Oral paper:

Learning Dynamics of LLM Finetuning!

(0/12)

From the intro to a paper I'm reading:

"Theorem 3 is proven in Section 4 and Theorem 4 is proven in Section 3."

😔

Some Like It Hot - Waterpolo YouTube video by ThMasker

youtu.be/0zG8Je6NyYM

Bias Amplification in Language Model Evolution: An Iterated Learning Perspective With the widespread adoption of Large Language Models (LLMs), the prevalence of iterative interactions among these models is anticipated to increase. Notably, recent advancements in multi-round self-i...

If for some reason you’re *not* in Vancouver for NeurIPS, you can always read the paper I guess:

arxiv.org/abs/2404.04286

What happens when models, like LLMs, learn from each other across multiple generations?

@joshuaren.bsky.social + collaborators extend a cogsci framework called Bayesian iterated learning to think about, and avoid, pitfalls.

#NeurIPS2024 poster #3305, Friday 11-2

(Also, snag him for a postdoc 😇)

This analysis of how attention scores work in graph transformers was pretty surprising to me, and enabled the rest of this paper (and hopefully more!). Come check out the poster soon!

As in Tennessee, this ban does not apply to cis children, because the drugs themselves are not a safety hazard. The only thing at risk is a society’s discomfort with allowing young people to make their own decisions about their own bodies.

Graph Transformers (GTs) can handle long-range dependencies and resolve information bottlenecks, but they’re computationally expensive. Our new model, Spexphormer, helps scale them to much larger graphs – check it out at NeurIPS next week, or the preview here!
[1/13]
#NeurIPS2024

A schedule of the planned events at the Queer in AI workshop at NeurIPS 2024. It will be held in West Meeting Room 202-204 at the Vancouver Convention Centre on 11 December. Opening Talk at 10 AM. Speaker session on "Designing Technology for Gender Transition" at 10:30 AM. Panel on "Queer Creative Storytelling" at 11 AM. Panel on "AI and Data Governance" at 1 PM. Co-working Session on "Participatory AI and Data Governance 1" at 2 PM. Sponsor Networking Session at 3 PM. Co-working Session 2 at 3:30 PM. Closing talk on "Queer Producing in Latin America" at 4:30 PM.

🌈🌈We have an incredible day planned for you at our #QueerInAI workshop at #NeurIPS2024 in Vancouver on December 11! You can join us in person or virtually via live-stream! Find out more about our events and our lovely speakers at queerinai.com/neurips-2024!! ✨✨

en.wikipedia.org/wiki/Poisson... has an upper and lower bound on the probability but not an exact value….

Oh, guess I should add that this is from arxiv.org/abs/2409.06890

Screenshot from the linked paper, showing Proposition 4 which relates HSIC to the mutual information. The proof is based on Pjnsker’s inequality and the Bretagnolle-Hiber bound, with a pointer to Canonne (2023).

I used the same note recently – it’s a good note!

bsky.app/profile/quai...

Well, I don’t know what else I was expecting.

The main reason I wrote arXiv-collector was to use biblatex cleanly, which can otherwise be a huge pain. If you don’t use biblatex, either is fine; each has some features the other doesn’t (though I prefer my way obviously :p).

GitHub - djsutherland/arxiv-collector: A little Python script to collect LaTeX sources for upload to the arXiv. A little Python script to collect LaTeX sources for upload to the arXiv. - djsutherland/arxiv-collector

Periodic reminder: submitting on #arXiv is great! But the source is available: good per se, but remember to remove LaTeX comments you don't want to be made public.

Thankfully, there are tools that do that (and more) easily and automatically for you! E.g., arXiv-collector:
github.com/djsutherland...