Update: I left Meta yesterday.
Working on
@OpenCatalyst
at FAIR has been one of my most cherished career journeys. The team's really strong, with ambitious step-change research in the pipeline. I'm rooting for them!
As for what's next, excited to build something new! Stay tuned
Excited to share that I successfully defended my PhD day before yesterday!
Huuuge thanks to
@DhruvBatraDB
and
@deviparikh
for being amazing mentors on this journey!
New preprint! TarMAC, a simple and effective architecture for learning targeted communication protocols in multi-agent RL. Leads to intuitive cooperation strategies on diverse environments.
w/
@theo_gervet
, Joshua,
@DhruvBatraDB
,
@deviparikh
, Mike, Joelle
We're looking for research interns to join us at FAIR as part of . If you're a PhD student interested in graph neural networks for modeling atomic relaxations, that will then be used to explore catalysts for renewable energy storage, please reach out to me!
Introducing —
@deviparikh
and my attempt at reducing some of the information asymmetry in industry job offers in AI.
Consider using this as a reference for negotiations. Covers a tiny slice of companies in US atm but should get wider coverage over time.
Introducing AI Paygrades ()! Statistics of industry offers for AI jobs. The goal is to reduce information assymetry so candidates can make informed decisions and negotiate better. Submit your information and spread the word! With
@abhshkdz
.
Excited for the first major release of , result of a wonderful ongoing collaboration b/w FAIR and CMU.
Our goal is to use ML to accelerate the search for low-cost catalysts that can drive reactions converting renewable energy to storable forms.
(1/8)
Panoptic Segmentation, neat work from FAIR / Heidelberg University on unifying instance and semantic segmentation, and a new intuitive evaluation metric for the task, with human studies + machine baselines
Astronauts flying around in space at 27k+ km/hr, launched on a reusable rocket, all being globally livestreamed from multiple camera angles with commentary. What a time. 🌎🚀
Happy to share that
@paperswithcode
will now be running !
It's been fun to see this random side project evolve into something so many from the AI community use (20k-30k active users / month) and help maintain (600+ forks, 300+ PRs)!
🤝 Papers with Code is now helping to maintain .
AI Conference Deadlines makes it easy to find and follow conference deadlines in areas such as machine learning and NLP. Thanks to
@abhshkdz
for this great initiative.
How much environment knowledge can an agent learn purely through self-supervised egocentric prediction?
We measure this by asking it questions (in English)!
#ICML2020
paper from my internship at
@DeepMind
. [1/4]
Honored to be selected as a runner-up for the 2020 AAAI/ACM SIGAI Dissertation Award! Huge thanks to the committee, my mentors and collaborators. I had the time of my life during my PhD, and often joke about how I might gladly do another :)
Congratulations to David Abel (A Theory of Abstraction in Reinforcement Learning) and Abhishek Das (Building Agents that Can See, Talk, and Act) for Honorable Mentions for the 2020 AAAI/ACM SIGAI Dissertation Award.
#AAAI2022
New preprint! TarMAC, a simple and effective architecture for learning targeted communication protocols in multi-agent RL. Leads to intuitive cooperation strategies on diverse environments.
w/
@theo_gervet
, Joshua,
@DhruvBatraDB
,
@deviparikh
, Mike, Joelle
ML research tool idea: self-updating academic benchmarks; that auto-scrapes arxiv, parses pdfs to build a unified table of results (on say CIFAR, Imagenet, COCO, etc.). Easily exportable to latex, csv.
Should make it v easy to answer “What is the current state-of-the-art on X?”.
Just came across this year's Google Summer of Code (
@gsoc
) stats:
Heartening to see my alma mater,
@iitroorkee
, top this year's list of schools with the most accepted students! 😃
Kudos to
@sdslabs
and everyone involved! 💪💥
That one day of the year you spend all day updating your bashrc / zshrc, vimrc, tmuxrc, gitconfig, sublime plugins, latex snippets, themes, fonts, colors at the cusp of transforming into an unstoppable maximally productive force of nature :)
Super excited to share AdsorbML (from
@OpenCatalyst
):
We show that machine learning can speed up density functional theory calculations for computing adsorption energies by * 1300x * at 85% accuracy (and by up to * 4000x * at 75% accuracy).
🧵
Yesterday we released data + code for our CVPR 2018 paper on EmbodiedQA. Train embodied agents for navigation and question answering in environments!
Data:
Code:
We're looking for research interns to come work with us as part of at FAIR. Applicants need to be PhD students, ideally interested in graph neural networks / ML for climate change, and available for 12-16 weeks over summer 2021. Get in touch with me!
Acceptance vs. paper id for CVPR/ICCV. CVPR18 looks a lot more evenly distributed than last few cycles (could be due to reviewer fatigue / worse paper quality <—> later registrations / changes in policy that I don't know of).
I'm at
#NeurIPS2022
!
Happy to chat about
@OpenCatalyst
and FAIR. We're also looking for research interns and residents for 2023, so please get in touch if you're interested.
New state-of-the-art on Visual Dialog! Multi-step attention architecture from Gan et al. / Microsoft. (Compares against published and unpublished work :)).
Slides for our
#acl2018
tutorial are now up on !
Tying together relevant ideas/works in a single coherent story took a fair bit of heavy-lifting. Was a fun exp. overall! We hope it was useful, and any feedback is very welcome :)
Excited to announce the 3rd
#NeurIPS2019
workshop on Visually-Grounded Interaction and Language (ViGIL)!
We have an exciting lineup of speakers:
— Linda Smith
— Josh Tenenbaum
— Jay McClelland
— Lisa Anne Hendricks
— Jesse Thomason
(more to be added)
* Visual Dialog challenge update *
~300 submissions from ~40 teams so far!
~One month to go!
$10k in GCP credits to be won!
Winners TBA at CVPR.
Pytorch starter code now supports both generative & discriminative models:
now has ~120+ verified offers! Huge thanks to Rora () for contributing ~75 of them (with permission)
Those on the AI job market this year -- consider adding your data if you're comfortable. More data on this = lesser info asymmetry
Did the Grand Canyon rim-to-rim this past weekend. ~24 miles total, -5k ft down, +6k ft up, 100+ temperatures. Probably the most physically challenging hike I’ve done, but gorgeous views all throughout!
Attending
#NeurIPS2019
till Friday!
Come say hi if you want to discuss grounded language learning or climate change or how to survive a PhD or just want to go get some sushi :)
New architecture for embodied question answering: decomposing navigation into high-level, interpretable subgoal prediction and low-level control
Presenting as a spotlight at
#corl2018
. Work done with Georgia Gkioxari,
@stefmlee
,
@deviparikh
,
@DhruvBatraDB
Excited to share a couple of updates on the Open Catalyst Project!
We've updated our paper with DimeNet++ (Klicpera et al.) results: . This is now the best-performing model for ML-driven relaxations on this task (still far from practical use tho).
(1/3)
EquiformerV2 was accepted at
#ICLR2024
!
EquiformerV2 is the result of (Yi-Lun's) careful empirical analysis of how to incorporate efficient equivariant convolutions into transformers, resulting in an architecture that scales surprisingly well across atomic datasets.
🧵(1/7) Excited to share that our work, EquiformerV2, has been accepted to
#ICLR2024
. EquiformerV2 is the state-of-the-art on large-scale atomistic benchmarks -- OC20, OC22, AdsorbML, and ODAC23.
Joint work with
@bwood_m
,
@abhshkdz
from
@OpenCatalyst
and
@tesssmidt
Paper:
New
#ICLR2022
paper on Graph Parallelism () — a distributed training method for GNNs with higher-order interactions (triplets / quadruplets) common in modeling atomic systems.
Work led by
@anuroopsriram
as part of
@OpenCatalyst
.
Just finished reading Elephant in the Brain by Simler and Hanson.
By far one of the most (almost depressingly) influential books I've read in a while. Irrespective of whether you agree / disagree, like / hate what's said, helps build an accurate world model. Highly recommended.
Interesting results from this year's Stack Overflow >100k developer survey
— TF is most loved, torch/pytorch 3rd (against React / Node / Django etc.)
— More excited about possibilities than potential dangers of AI
— Sufficiently concerned about AI biases
In Montréal for a summer internship; first thought — feels good to be back in metric-system-land and not have do mental math to go from pixels to understanding :)
Recently watched Joker and absolutely loved it.
Dark, intense, thought-provoking. Flips the typical villain narrative on its head. Ferocious performance by Joaquin Phoenix 🙏🏽
Watch if you haven't already. Highly recommended.
23 extra minutes of daily commute has the same effect on happiness as a ~19% reduction in income! :o (controlled for a lot of ~obvious response biases)
From Stress That Doesn't Pay: The Commuting Paradox (Stutzer and Frey, 2004)
Acceptance vs. paper id for CVPR/ICCV. CVPR18 looks a lot more evenly distributed than last few cycles (could be due to reviewer fatigue / worse paper quality <—> later registrations / changes in policy that I don't know of).
Join
@panderson_me
, Qi Wu and me for our
@acl2018
tutorial on Connecting Language and Vision to Actions tomorrow!
July 15, 9:30 AM — 12:30 PM, Room 218, MCEC.
Peter made a cool viz showing some of the topics we'll be covering :)
My
#ICLR2020
peer review review: Kudos to the one (and only) high-quality reviewer out of six across two papers.
(Also the only one that recommended rejection ¯\_(ツ)_/¯)
Roommates introduced me to Gaia Project this weekend, and we ended up playing a marathon 7-hour game!
Way too many rules to keep track of, but amazing game design and space colonization theme; it’s now overtaken Terraforming Mars as my new favorite :)