Co-founder and Co-CEO, Yutori
Bio
I am the Co-founder and Co-CEO of Yutori, where we’re building state-of-the-art web agents – AI agents that can reliably take actions and execute tasks on the web.
Previously, I was a Research Scientist at Fundamental AI Research (FAIR) at Meta working on deep neural networks and its applications in climate change, specifically focusing on electrocatalyst discovery for renewable energy storage as part of the Open Catalyst Project.
Before that, I was a Computer Science PhD student at Georgia Tech, advised by Dhruv Batra, and working closely with Devi Parikh, where I focused on developing artificial agents that can see (computer vision), talk (language modeling), and act (reinforcement learning).
2011 - 2015
Summer 2015
2015 - 2016
2017 - 2020
S2017, W2018, S2018
Winter 2019
Summer 2019
2020-2024
During my PhD, I interned thrice at Facebook AI Research — Summer 2017 and Spring 2018 at Menlo Park, working with Georgia Gkioxari, Devi Parikh and Dhruv Batra on training embodied agents for navigation and question-answering in simulated environments (see embodiedqa.org), and Summer 2018 at Montréal, working with Mike Rabbat and Joelle Pineau on communication protocols in multi-agent reinforcement learning. In 2019, I interned at DeepMind in London working on grounded language learning with Felix Hill, Laura Rimell, and Stephen Clark, and at Tesla Autopilot in Palo Alto working on differentiable neural architecture search with Andrej Karpathy.
My PhD research was supported by fellowships from Facebook, Adobe, and Snap.
Prior to joining grad school, I worked on neural coding in zebrafish tectum as an intern under Prof. Geoffrey Goodhill and Lilach Avitan at the Goodhill Lab, Queensland Brain Institute.
I got my Bachelor’s at IIT Roorkee in 2015. During my undergrad, I took part in Google Summer of Code (2013 and 2014), won several competitions (Yahoo! HackU!, Microsoft Code.Fun.Do., Deloitte CCTC 2013 and 2014), and owe most of my programming/tinkering bent to SDSLabs.
On the side, I built aideadlin.es (countdowns to a bunch of CV/NLP/ML/AI conference deadlines) and aipaygrad.es (statistics of industry job offers in AI), neural-vqa and its extension neural-vqa-attention, HackFlowy, graf, Erdős, etc. I also occasionally dabble in generative art. I like this map tracking the places I’ve been to. Blog posts from a previous life.
Talks and Interviews
Publications
UMA: A Family of Universal Models for Atoms
Generalizing Denoising to Non-Equilibrium Structures Improves Equivariant Force Fields
TMLR 2024
Paper
The Open DAC 2023 Dataset and Challenges for Sorbent Discovery in Direct Air Capture
ACS Central Science 2024
Paper
Code
Website
EquiformerV2: Improved Equivariant Transformer for Scaling to Higher-Degree Representations
AdsorbML: Accelerating Adsorption Energy Calculations with Machine Learning
npj Computational Materials 2023
Paper
Code
PIRLNav: Pretraining with Imitation and RL Finetuning for ObjectNav
CVPR 2023, RRL workshop at ICLR 2023
Paper
Code
Website
The Open Catalyst 2022 (OC22) Dataset and Challenges for Oxide Electrocatalysis
ACS Catalysis 2023
Paper
Code
Dataset
GemNet-OC: Developing Graph Neural Networks for Large and Diverse Molecular Simulation Datasets
Spherical Channels for Modeling Atomic Interactions
Open Challenges in Developing Generalizable Large Scale Machine Learning Models for Catalyst Discovery
ACS Catalysis (Perspective) 2022
Paper
Transfer learning using attentions across atomic systems with graph neural networks (TAAG)
The Journal of Chemical Physics 2022
Paper
Code
Habitat-Web: Learning Embodied Object-Search Strategies from Human Demonstrations at Scale
CVPR 2022
Paper
Code
Website
Presentation video
Towards Training Billion Parameter Graph Neural Networks for Atomic Simulations
Rotation Invariant Graph Neural Networks using Spin Convolutions
Automated Video Description for Blind and Low Vision Users
CHI EA 2021
Paper
Auxiliary Tasks and Exploration Enable ObjectNav
ForceNet: A Graph Neural Network for Large-Scale Quantum Calculations
ICLR 2021 Deep Learning for Simulation Workshop
Paper
opencatalystproject.org
Presentation video
The Open Catalyst 2020 (OC20) Dataset and Community Challenges
ACS Catalysis 2021
Paper
Code
Dataset
opencatalystproject.org
An Introduction to Electrocatalyst Design using Machine Learning for Renewable Energy Storage
Auxiliary Tasks Speed Up Learning PointGoal Navigation
Large-scale Pretraining for Visual Dialog: A Simple State-of-the-Art Baseline
Building agents that can see, talk, and act
PhD Thesis
AAAI/ACM SIGAI Doctoral Dissertation Award, Runner-up
Georgia Tech Sigma Xi Best PhD Thesis Award
Georgia Tech College of Computing Dissertation Award
Probing Emergent Semantics in Predictive Agents via Question Answering
ICML 2020
Paper
Presentation video
Slides
Feel The Music: Automatically Generating A Dance For An Input Song
IR-VIC: Unsupervised Discovery of Sub-goals for Transfer in RL
IJCAI-PRICAI 2020, ICLR 2019 Task-Agnostic RL Workshop
Paper
Improving Generative Visual Dialog by Answering Diverse Questions
TarMAC: Targeted Multi-Agent Communication
Embodied Question Answering in Photorealistic Environments with Point Clouds
CVPR 2019 (Oral)
Paper
Audio-Visual Scene-Aware Dialog
CVPR 2019
Paper
Code
video-dialog.com
End-to-end Audio Visual Scene-Aware Dialog Using Multimodal Attention-based Video Features
ICASSP 2019
Paper
video-dialog.com
Neural Modular Control for Embodied Question Answering
CoRL 2018 (Spotlight)
Paper
embodiedqa.org
Presentation video
Slides
Embodied Question Answering
CVPR 2018 (Oral)
Paper
embodiedqa.org
Code
Presentation video
Slides
Evaluating Visual Conversational Agents via Cooperative Human-AI Games
Learning Cooperative Visual Dialog Agents with Deep Reinforcement Learning
ICCV 2017 (Oral)
Paper
Code
Presentation video
Slides
Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization
IJCV 2019, ICCV 2017, NIPS 2016 Interpretable ML for Complex Systems Workshop
Paper
Code
Demo
Visual Dialog
PAMI 2018, CVPR 2017 (Spotlight)
Paper
Code
visualdialog.org
AMT chat interface
Demo
Presentation video
Slides
Human Attention in Visual Question Answering: Do Humans and Deep Networks Look at the Same Regions?
CVIU 2017, EMNLP 2016, ICML 2016 Workshop on Visualization for Deep Learning
Paper
Project+Dataset
neural-vqa-attention
Side projects
aipaygrad.es
aipaygrad.es provides statistics of industry job offers in Artificial Intelligence (AI).
All data is anonymous, cross-verified against offer letters and will
hopefully reduce information asymmetry.
aideadlin.es
aideadlin.es is a webpage to keep track of CV/NLP/ML/AI conference deadlines. It's hosted on GitHub, and countdowns are automatically updated via pull requests to the data file in the repo.
neural-vqa-attention
Torch implementation of an attention-based visual question answering model (Yang et al., CVPR16).
The model looks at an image, reads a question, and comes up with an answer to the question and a heatmap of where it looked in the image to answer it.
Some results here.
neural-vqa
neural-vqa is an efficient, GPU-based Torch implementation of the visual question answering model from the NIPS 2015 paper 'Exploring Models and Data for Image Question Answering' by Ren et al.
Erdős
Erdős by SDSLabs is a competitive math learning platform, similar in spirit to Project Euler, albeit more feature-packed (support for holding competitions, has a social layer) and prettier.
graf
graf plots pretty git contribution bar graphs in the terminal.
gem install graf to install.
HackFlowy
Clone of WorkFlowy.com, a beautiful, list-based note-taking website that has a 500-item monthly limit on the free tier :-(. This project is an open-source clone of WorkFlowy. "Make lists. Not war." :-)
AirMaps
AirMaps was a fun hackathon project that lets users navigate through Google Earth with gestures and speech commands using a Kinect sensor. It was the winning entry in Microsoft Code.Fun.Do.
HackView
Another fun hackathon-winning project built during Yahoo! HackU! 2012 that involves webRTC-based P2P video chat, and was faster than any other video chat provider (at the time, before Google launched Hangouts).
8tracks-downloader
Ugly-looking, but super-effective bash script for downloading entire playlists from 8tracks. (Still works as of 10/2016).