Machine learning lab at Columbia University. Probabilistic modeling and approximate inference, embeddings, Bayesian deep learning, and recommendation systems.
We have a new paper on topic models!
We introduce Mechanistic Topic Models (MTMs). MTMs model SAE feature counts rather than words or plain embeddings, combining the benefits of LLM-based and probabilistic topic models.
Paper (published in TACL) : arxiv.org/abs/2507.23220
We show that MTMs are better than traditional and neural baselines at finding salient and abstract themes, while costing less than LLM-based methods.
Moreover, working with SAE features allows you to construct topic steering vectors to steer LLM generation toward MTM topics.
1/ Two new preprints on OOD generalization.
Shared lesson: the goal should not always be to find what is invariant.
For good prediction in new environments, it can be better to model environment variation explicitly, then marginalize it out.
arXiv:2604.26128
arXiv:2606.05365
I am on the academic job market this year! My research advances probabilistic machine learning and its application to biology. I'm looking for faculty positions in stats/CS/applied math and in bio departments.
Academic website: eweinstein.github.io/
New paper from @blei_lab member @EliWeinstein6 on manufacturing samples from generative protein models. The idea is to approximate a complex distribution with simpler ones that are easy to sample from in the real world, using stochastic chemical reactions.
Excited to announce that ML-NYC is back this semester. Our first speaker is Bin Yu from UC Berkeley on Monday Sept 23rd 4pm. Register here: eventbrite.com/e/ml-nyc-spea…
We're excited to welcome Daniel Lee as our next speaker on Wednesday September 20th at 4pm!
The event will be followed by a happy hour hosted by the Flatiron Institute
Register here: eventbrite.com/e/ml-nyc-spea…
Belated updates:
1- I defended my PhD in May! Big thank you to my committee and especially to my advisor, @blei_lab.
2- I'm excited to start a postdoc at Harvard! Very grateful to @harvard_data for the opportunity.