Pinned
eating very spicy food is a great way to stop dwelling about something
Mark Goldstein
1,609 posts
Mark Goldstein
@marikgoldstein
absolutely discontinuous with respect to Lebesgue measure. research fellow at the flatiron institute in nyc @FlatironInst
- diffusion models are just a ploy by CS PhDs to get their departments to pay for them to learn stochastic calculus so that they can get finance jobs
- I have defended! thanks Rajesh Ranganath for mentoring me over these years. thanks Eric Vanden-Eijnden as well. thanks to the committee (Rajesh, Eric, Kyunghyun Cho, Rob Fergus, and Arthur Gretton). thanks to collaborators, friends, family, loved ones and espresso
- I'm mark, I like diffusion models, AI/ML for science + health, and espresso. Let me know if you'd like to chat at neurips! and for those hiring: i'm finishing my phd at NYU this year and am searching for jobs starting Fall 2025 (postdoc or industry) pictured: an old friend
- finishing up undergrad (or did so within past 3 years) and want to transition to CS + ML + intelligence research? Want to do a PhD in future? Some places have fully funded post-bacs! here's one at Harvard Kempner kempnerinstitute.harvard.edu/education/post…
- I cant stand that "multi-modal" means both "several kinds of data" and "a distribution with multiple modes" (density local optimas)
- Replying to @TheGregYangI met you right after this. I had just switched to CS after doing music undergrad. I remember asking you which classes to take; you said, doesn’t matter, study yourself, at home, in the dining hall etc… best advice (you did help me w/ a complexity problem that night tho)
- if people are just tokenizing everything and throwing transformers at it, what are some perspectives on why non-autoregressive models remain important? I have a few thoughts but would love to hear others' opinions on this
- Really cool work! Great gains from just a few particles and some smart choices of potentials!A General Framework for Inference-time Scaling and Steering of Diffusion Models Introduces Feynman-Kac steering, an inference-time steering framework for sampling diffusion models guided by a reward function. It generates multiple samples (particles) like best-of-n (importance
- what are some exciting places to do a postdoc in generative models / diffusion + applications in the health and sciences?
- New priors for your flows! We show how to build conditional models that flow directly from the conditioned data, rather than from scratch. Stochastic Interpolants with data-dependent couplings: arxiv.org/abs/2310.03725 w/ the incredible @msalbergo @nmboffi R. Ranganath @evdende2
GIF
🥳New paper thread🥳: We show how to build flows/diffusions between *dependent distributions*. The benefits seem really cool! Stochastic Interpolants with data-dependent couplings: arxiv.org/abs/2310.03725, big thanks @marikgoldstein @nmboffi Rajesh Ranganath, Eric Vanden-Eijnden - glad to share our #ICLR2023 work on extending diffusions to a broad set of new forward noising processes, derivation-free! arxiv.org/abs/2302.07261 We find that choosing the process for a fixed architecture can account for as much BPD improvement as a 3x increase in model size1/ We are excited to share new work on diffusion-based generative models (DBGMs)! This work introduces a large class of noising processes (diffusions) that are easy to use and shows the benefits of building DBGMs with them. Joint work with @marikgoldstein Rajesh Ranganath.
- cold rainy days in nyc happen when someone uses the portal between deepmind london and deepmind new york
