Pinned
Gergely Neu
3,722 posts
Gergely Neu
@neu_rips
ML theory nerd & AI non-enthusiast
- if you want a professor to reply immediately to your email, send it during a department meeting
- me on a normal day: "ugh i'm so tired of this AI hype" me when writing grant proposals:
- i've been submitting my work to conferences for more than 10 years, fighting to make each deadline. i have tried many strategies to cope with the anxiety and make sure that i don't have to work insane amounts of hours on the days before the deadline. here's what i learned: 👇🧵
- RL folks, meet your newest friend: THE LOGISTIC BELLMAN ERROR A convex loss function derived from first principles of MDP theory that leads to practical RL algorithms that can be implemented without *any* approximation of the theory. Preprint: arxiv.org/abs/2010.11151 🧵👇 1/18
- MYSTERY: why does SGD generalize well? My new paper may provide some answers: a perturbation analysis identifies three factors that contribute to good generalization. "Flatness" is one of them. (And yes, the analysis is information-theoretic.) arxiv.org/abs/2102.00931 1/
- not sure i agree. my alternative take: in computer science, 2% of scientists get 98% of the attention. attracting attention and doing useful work are definitely not equivalent. there's so much amazing work out there that is not getting the recognition it deserves.In computer science, 2% of scientists do 98% of useful research.
- bullshit ML term of the day: "information-theoretic" most common meaning: "we take a logarithm somewhere in the algorithm / analysis" please give it up already, big words don't make you look smarter.
- thinking about giving myself a 12-month deadline extension for @NeurIPSConf
- my past experience with corporate party invitations at NeurIPS: 2010-2013, PhD student @ obscure univ ---> 0 invites 2014-2015, postdoc @ super famous lab ---> 346 invites 2016-, junior faculty @ obscure-ish univ ---> 0 invites propose an OPEN social event & fight this system!NeurIPS 2019 Call for socials. nips.cc/Conferences/20… We are very happy to inaugurate social events at the coming NeurIPS 2019. Given the steadily increasing amount of attendees of NeurIPS, it's becoming more and more difficult to meet colleagues with similar interests ...
- who decided to call it "generator network" instead of "deceptron"??
- OK, time for some tweets about distances between Markov chains! Actually this is about a preprint we've just posted on arxiv with Sergio Calo, Anders Jonsson, Ludovic Schwartz & Javier Segovia-Aguas. FFO optimal transport & bisimulation. Let's dig in! arxiv.org/abs/2406.04056 1/n
