Divyansh Garg

Intro

Hi, I'm Div. I am a computer scientist and AI researcher. I am on a leave of absence from a CS PhD at Stanford University. I am building MultiOn - a Personal AI Agent startup with the mission of building general-purpose AI similar to JARVIS to do your digital tasks and bring it to everyday consumer use.

I am passionate about all things Deep Learning as well as its applications to Robotics and Vision. My recent focuses have been on building systems that are generally intelligent and able to reason, designing learning algorithms that can learn as efficiently as humans and making Reinforcement Learning work in the real world.

I created and taught the first course on Transformers at Stanford — CS 25: Transformers United — discussing the latest breakthroughs and broad implications of Transformers in AI. The class invites people at the forefront of Transformers research in various fields to spark cross-collaborative research, including eminent speakers like Prof. Geoffrey Hinton. Our lectures have received an overwhelming public reception with over 1M views on Youtube.

I have given invited talks at OpenAI, DeepMind and Apple on my research. I recently worked on a new RL algorithm — Extreme Q-Learning — that builds on a 2000 Nobel prize work in Economics to solve a previously assumed intractable problem and reaches state-of-the-art in Offline RL. My past work — IQ-Learn — introduces a novel imitation learning framework to have AI agents learn from observing videos of humans, creating agents that can play video games like Atari at human performance and won the #1 place in NeurIPS '21 Minecraft AI competition. The work received media coverage from Stanford HAI.

In my undergrad at Cornell, I created the first working camera-based self-driving car system and published 4 major conference papers over 2 years in 3D Vision. My work was widely covered in the media: Forbes, The Robot Report, Gizmodo.

I have been fortunate to be mentored by Ian Goodfellow (and was his first intern at Apple).

In a past life, I was a child prodigy in Physics and won medals in a few International Olympiads.

Experience

Apple SPG

March '20 - Sept '20

I worked as a research intern in Apple Special Projects Group. Directly supervised by renowned researcher Ian Goodfellow. Researched on RL, Inverse RL and Generative Modeling.

Google AI

Summer '19

I interned in Mountain View on the Google Machine Perception team. I designed ML models to solve real-time computer vision problems for AR devices.

CV Research

Aug '18 - May '20

I did Computer Vision research with Prof. Kilian Q. Weinberger and Prof. Bharath Hariharan . My focus was on camera-only depth estimation and 3D object detection. I created a state-of-the-art model for stereo-only 3D object detection.

Uber ATG

Summer '18

I interned in Uber's self-driving car unit in Pittsburgh. I worked on the Perception team and improved the autonomous vehicle’s 3D object detection system.

Cornell Mars Rover

Sept '17 - Dec '18

I worked on a Cornell project team to build a rover to compete in the University Rover Challenge in Utah. I built autonomous systems using Computer Vision to achieve object detection and improved on the navigational abilities of the rover.

COMAKE

Aug '17 - Dec '17

I worked as a ML Engineer to create a smart file browser with file analysis and context-based workflow management abilities. I designed and implemented a system to recommend related files and improve the user’s work experience.

Publications

Extreme Q-Learning: MaxEnt RL without Entropy

ICLR 2023 (Oral)

Divyansh Garg*, Joey Hejna*, Matthieu Geist, Stefano Ermon

Paper

tl;dr Introduce Gumbel Regression to learn maximal values in RL without needing to sample from a policy and reaches SOTA performance on offline RL.