Seth Neel
1.1K posts
@SethInternet
📍something new prev: rs @googleai (gemini data), asst professor @harvard (biz + cs affil), co-founder @welligence, PhD @penn
Machine unlearning ("removing" training data from a trained ML model) is a hard, important problem. Datamodel Matching (DMM): a new unlearning paradigm with strong empirical performance! w/ @kris_georgiev1 @RoyRinberg @smsampark @shivamg_13 @aleks_madry @SethInternet (1/4)
.@joemelko argues that data curation should be treated as an optimization problem, not guesswork. he walks us through how to learn ... how to learn.
TIL that the asymptotically unbiased estimator produced by MCMC can be turned into an unbiased estimator using only a *finite* number of iterations! The trick is v neat: Turn an infinite limit into a telescoping sum, then arrange to have all terms after some step t cancel out.
New paper alert!🚨 What do LLM reasoning, diffusions, & jailbreaks have in common? 🤔 All exhibit critical windows📈--a sudden formation of distinctive features during sampling, e.g. correctness or toxicity. We present a unifying theory of critical windows for diffusion & LLMs.
Our mission is to develop the best technology for interfacing with the brain to improve people’s lives.
