Papers of the day   All papers

RTFM: GENERALISING TO NOVEL ENVIRONMENT DYNAMICS VIA READING

Comments

Tim Rocktäschel: Fantastic work by Victor Zhong (@hllo_wrld) during his internship at @FacebookAI Research London. Procedurally generating environment dynamics and their textual descriptions forces the agent to perform multi-hop reading comprehension. Paper: https://arxiv.org/abs/1910.08210 w/ @egrefen https://t.co/I3XFqQTkQx

6 replies, 185 likes


Victor Zhong: Super excited to share our @iclr_conf 2020 work RTFM: Generalising to Novel Environment Dynamics via Reading (https://arxiv.org/abs/1910.08210)! Joint work w/ @_rockt and @egrefen at @facebookai research London on RL policies that generalise to new envs via reading! 1/3 👇 https://t.co/n9wL4m6gsD

2 replies, 151 likes


Edward Grefenstette: One thing to highlight is the FiLM² layer introduced in §4.1 of the paper, which is a particularly simple-yet-powerful way of cross-conditioning from 2+ modalities. We did text/vision but in principle this works for anything… @pytorch code for this layer: https://github.com/facebookresearch/RTFM/blob/master/model/paper_txt2pi.py#L12 https://t.co/pIEOqalSSv

1 replies, 138 likes


hardmaru: RTFM: Generalising to New Environment Dynamics via Reading They propose “Read to Fight Monsters,” a problem in which the agent must jointly reason over a language goal, relevant dynamics described in a document, and environment observations. open review: https://openreview.net/forum?id=SJgob6NKvH https://t.co/jViS1t1uM9

0 replies, 47 likes


Edward Grefenstette: The code for our RTFM task suite and text2π architecture (in @PyTorch) is now available at https://github.com/facebookresearch/RTFM! Great work by @hllo_wrld! Read the paper: https://arxiv.org/abs/1910.08210 and a blog post about the work: https://ai.facebook.com/blog/read-to-fight-monsters-using-rl-to-teach-agents-to-generalize-to-new-settings/ https://t.co/cZjIepYuCF

0 replies, 45 likes


Victor Zhong: I had a brilliant 😉 time interning at @facebookai research London - one of the best places for #nlproc + #ReinforcementLearning research. A huge thank you to @egrefen @_rockt and @riedelcastro for having me there!

1 replies, 44 likes


Tim Rocktäschel: Victor (@hllo_wrld) did outstanding work during his internship with @egrefen and me at @FacebookAI Research in London last summer. Really happy his paper will be presented at @iclr_conf 2020 in Addis Ababa! See below for a short summary and stay tuned for the open-source release.

0 replies, 38 likes


Edward Grefenstette: Looking forward to giving a brand new talk (in that I just finished writing it) at @mldcmu/@LTIatCMU this afternoon! https://www.scs.cmu.edu/calendar/tue-2020-08-11-1100/special-machine-learning-language-technologies-institute-talk

0 replies, 33 likes


Tim Rocktäschel: Great starting point for Language in RL projects. We open-sourced the procedurally generated RTFM environment, a FiLM² layer implementation, as well as PyTorch RL agent code based on TorchBeast/IMPALA.

0 replies, 25 likes


Graham Neubig: Great to have you @egrefen! For those who couldn't make it, the interesting talk covered NetHack https://github.com/facebookresearch/nle AMIGo https://arxiv.org/abs/2006.12122 RTFM https://arxiv.org/abs/1910.08210 and more!

1 replies, 18 likes


Edward Grefenstette: Happening NOW 😃 https://iclr.cc/virtual/poster_SJgob6NKvH.html

0 replies, 13 likes


Edward Grefenstette: Check out this thread for a nice little summary of @hllo_wrld’s awesome internship project with @_rockt and me, done at @facebookai last summer. Upcoming at @iclr_conf 2020!

0 replies, 12 likes


Tim Rocktäschel: Join us in the #ICLR2020 poster session on Thursday 05:00 to 07:00 & 17:00 to 19:00 GMT https://iclr.cc/virtual/poster_SJgob6NKvH.html

0 replies, 12 likes


Victor Zhong: Our @iclr_conf poster sessions will be this Thursday 5-7am GMT / 10pm-12am PDT during Session 1 and 5-7pm GMT / 10pm-12am PDT during Session 4. More details here: https://iclr.cc/virtual/poster_SJgob6NKvH.html Hope to see you there!

0 replies, 10 likes


Edward Grefenstette: @facebookai @jelennal_ @CompSciOxford @PyTorch In @hllo_wrld's internship project, with @_rockt, we show that RL agents can zero-shot new settings when trained to exploit supporting documentation. Upcoming at ICLR 2020 (12/16) https://arxiv.org/abs/1910.08210 https://t.co/o21hqaDrJG

1 replies, 4 likes


Montreal.AI: RTFM: Generalising to Novel Environment Dynamics via Reading Zhong et al.: https://arxiv.org/abs/1910.08210 #ArtificialIntelligence #MachineLearning #ReinforcementLearning https://t.co/AQugV1PMXE

0 replies, 4 likes


cs.LG Papers: RTFM: Generalising to Novel Environment Dynamics via Reading. Victor Zhong, Tim Rocktäschel, and Edward Grefenstette http://arxiv.org/abs/1910.08210

1 replies, 1 likes


arXiv CS-CL: RTFM: Generalising to Novel Environment Dynamics via Reading http://arxiv.org/abs/1910.08210

0 replies, 1 likes


arXiv CS-CL: RTFM: Generalising to Novel Environment Dynamics via Reading http://arxiv.org/abs/1910.08210

0 replies, 1 likes


Brundage Bot: RTFM: Generalising to Novel Environment Dynamics via Reading. Victor Zhong, Tim Rocktäschel, and Edward Grefenstette http://arxiv.org/abs/1910.08210

1 replies, 0 likes


Content

Found on Oct 21 2019 at https://arxiv.org/pdf/1910.08210.pdf

PDF content of a computer science paper: RTFM: GENERALISING TO NOVEL ENVIRONMENT DYNAMICS VIA READING