user avatar
galvanize (gail weiss) πŸ’”πŸŽ—οΈ
@gail_w
Trying to get off of here, dms checked rarely. Better to send an email gailweiss.github.io .
Posts
  • Pinned
    user avatar
    Seems like as good a time as ever to make (and maintain) a thread of publications, to keep them easy to find. First is extraction of DFAs from recurrent neural networks, using lstar. Challenge was equivalence queries.
    We have a cool new algorithm for extracting automata from RNNs (LSTMs, GRUs..) Turns out that for many simple languages, RNNs actually learns quite large and weird DFAs that have many blind-spots which our algo discovers. arxiv.org/abs/1711.09576 (w/ Gail Weiss, @yahave )
  • user avatar
    Fuck me did you know you have a little ghoulie in there
    Drawing of Gordon Ramsay staring in horror at a little mushroom man, who is eating a rotten cucumber like a banana, as described in the quoted tweet
  • user avatar
    okay this has probably been done but
  • user avatar
    Roses are red RNNs process sequences But if you look closely You’ll find they have weaknesses New work with @yoavgo and @yahave on the practical power of different RNN architectures available now on arxiv, and soon at @acl2018! arxiv.org/abs/1805.04908
  • user avatar
    EXTREMELY excited to announce RASP, a programming language whose goal is to provide a computational model for transformers in much the same way that automata have served for RNNs. Work with @yoavgo and @yahave , accepted into ICML 2021. arxiv.org/abs/2106.06981
    Screenshot from β€œThinking like transformers”. On one side, a short code snippet computing histograms in RASP. On the other, a printout of the compiled RASP architecture for that program, showing the selection pattern of this program when applied to the input sequence β€œ$aabbaabb”.  Selection patterns are RASP’s abstraction of a transformer’s attention pattern. Alongside this selection pattern, a heatmap of the single attention pattern of a one-layer one-head transformer trained on the same task (histograms) and applied to the same sequence. The selection pattern and the heatmap bear a strong resemblance.
  • user avatar
    Not a β€œstrong publication record” at ICML, NeurIPS, ACL, etc just to get into an *internship* come on
  • user avatar
  • user avatar
  • user avatar
    We doubled down on our L* attack on RNNs and now have an alg for learning probabilistic deterministic finite automata from a given language model! The alg is a weighted adaptation of L*, & easily reconstructs the target from RNNs trained on small PDFAs. arxiv.org/abs/1910.13895
  • user avatar
    My first ever email (correctly) addressing me as doctor went straight to junk, and then I couldn't attend my graduation on account of being in an entirely different country, but: hypers and onlookers, I am now a ✨ doctor ✨! Thanks to @yoavgo and @yahave for bringing me here!
    Mild throwback photo with Yoav, me, and Eran
    The screaming AAAAAA bird but with a little party hat and a graduation gown and cap on the side ready to go
    Grad school email came in, they have everything they need!!! It’s doctoring time babey!!!
  • user avatar
    Presenting: a method for synthesising CFGs from RNNs! The idea is to generalise the hypothesis DFAs extracted from the RNN by L-star, by identifying the sub-DFAs (patterns) that are added repeatedly between consecutive DFAs. arxiv.org/abs/2101.08200 great work with Danny Yellin
  • user avatar
    Why does my DFA look like some guy and a centaur squaring off under the watchful gaze of the Flying Spaghetti Monster
  • user avatar
    I'm βšͺ gay βšͺ straight πŸ”˜ A LUMBERJACK AND IM OKAY
    The giant redwood, the larch, the fir, the mighty Scots pine merriam-webster.com/words-at-play/…
  • user avatar
    Replying to @gail_w
    Simply give the kids money and a chance I promise you’ll be fine