Can SGD Select Good Fishermen? Local Convergence under Self-Selection Biases and Beyond We revisit the problem of estimating $k$ linear regressors with self-selection bias in $d$ dimensions with the maximum selection criterion, as introduced by Cherapanamjeri, Daskalakis, Ilyas, and Zamp...

"What makes a good fisherman as opposed to other professions?"
This question can be formulated as a k-linear regression problem with self-selection bias.

Alkis, @anaymehrotra.bsky.social, and I design faster local convergence algorithms for this problem:
arxiv.org/abs/2504.07133

(1/7)

Really big thanks to the organizers for the invitation & for putting together such a fun workshop.

My talk: simons.berkeley.edu/talks/andrew...

The paper: arxiv.org/abs/2503.13751

Joint work with @logn.bsky.social, Benjamin Chen, Axel Feldmann, Billy Moses, and @aleksmadry.bsky.social

Had a great time @simonsinstitute.bsky.social last week talking about new & upcoming work on meta-optimization of ML training

tl;dr: we show how to compute gradients *through* the training process & use them to optimize training. Immediate big gains on data selection, poisoning, attribution & more!

We'd love to hear your feedback if you attended the ATTRIB workshop at @neuripsconf.bsky.social 2024!

Please consider taking 2-3 min to fill out this anonymous form: forms.gle/JzGebsx9haD5...

Thank you!🙏

After another very lively poster session, our final talk of the day from @coallaoh.bsky.social - who is talking about the interactions between ML, attribution, and humans!

Our second-last talk of the day - Robert Geirhos on “how do we make attribution easy?”

One great poster session (and lunch) later - Baharan Mirzasoleiman on data selection for large language models!

After some amazing contributed talks, we now have a panel moderated by @sadhika.bsky.social - with @coallaoh.bsky.social Baharan Mirzasoleiman and Robert Geirhos!

Next up, @sanmikoyejo.bsky.social on predicting downstream properties of language models!

Our first talk of the day @ ATTRIB 2024 (Rm 205-207): @surbhigoel.bsky.social on attributing model behavior using synthetic data!

Giving a talk tomorrow at #NeurIPS2024 on the exciting topic of explainability!

ATTRIB 2024 WorkshopConference Schedule

At NeurIPS? Come by the 2nd workshop on Attributing Model Behavior at Scale (ATTRIB)!

Meeting Rm 205-207 @ 9am - amazing talks by @surbhigoel.bsky.social @sanmikoyejo.bsky.social Baharan Mirzasoleiman, Robert Geirhos, @coallaoh.bsky.social + exciting contributed talks!

Details: attrib-workshop.cc

Smoothed Analysis of Algorithms: Why the Simplex Algorithm Usually Takes Polynomial Time We introduce the smoothed analysis of algorithms, which is a hybrid of the worst-case and average-case analysis of algorithms. In smoothed analysis, we measure the maximum over inputs of the expected ...

You might be looking for smoothed analysis (en.wikipedia.org/wiki/Smoothe...)? Kind of interpolates between worst and average-case: no distribution over problem instances you have to specify but ignores "brittle" worst-case instances. Explains, eg, simplex algorithm (paper: arxiv.org/abs/cs/0111050)

I am recruiting PhD students at Duke!

Please apply to Duke CS or CBB if you are interested in developing new methods and paradigms for NLP/LLMs in healthcare.
For details, see here: monicaagrawal.com/home/researc....

Primarily written for the Operations market, but folks may find this guide I wrote for the job market: gargnikhil.com/files/Nikhil...