Vedant Misra

About me

I am a researcher, engineer, and entrepreneur passionate about advancing science and technology in order to develop new kinds of useful systems.

My background is in developing and shipping high-value automated systems by leveraging and advancing the state of the art in deep learning. My objective is to understand the computational principles of intelligence, and use them to rapidly accelerate social and technical progress.

Background

I work on artificial intelligence research at Google Research, focused on developing systems that can perform reasoning.

Previously, I worked with Ilya Sutskever to lead the Reasoning, Algorithms, and Multimodal teams at OpenAI, focused on developing methods for unlocking new capabilities that we believe to be on the path to artificial general intelligence.

Before that, I led the Machine Learning Labs team at HubSpot, focused on developing new products with text, images, and audio by leveraging large neural networks and human feedback.

Prior to that I was founder and CEO of Kemvi, which was acquired by HubSpot. Our vision was to accelerate the topline growth of businesses by using machine learning to understand and personalize language at scale.

I studied physics, mathematics, and economics at Columbia University as a Rabi Scholar, worked on black hole research at the Institute for Strings, Cosmology, and Astroparticle physics at Columbia, and studied complex systems at NECSI, an independent think tank focused on applying methods from physics and math to societal problems.

Projects

• Codex: We fine-tuned a GPT language model on publicly available code from GitHub, and studied its Python code-writing capabilities.

• Grokking: We studied generalization of neural networks on small algorithmically generated datasets and showed that in some situations neural networks learn through a process of “grokking” a pattern in the data, improving generalization performance from random chance level to perfect generalization, and that this improvement in generalization can happen well past the point of overfitting. (YouTube / Yannic Kilcher)

• Generative multimodal models: We released DALL-E, a system for generating images from text that was covered by CNBC, Engadget, the MIT Technology Review, TechCrunch, VentureBeat, New Scientist, and others, and developed by Aditya Ramesh.

• Medicine: We reviewed the use of artificial intelligence to enhance the detection and treatment of glaucoma.

• Language generation: We patented our work on training models to automatically generate personalized messages and studied how to extract information from such models via black box attacks.

• Theoretical astrophysics: We used non-linear optimization to characterize orbital trajectories near exotic black holes.

• Statistical finance: We used methods from quantitative finance to advise the SEC and House Financial Services Committee on policy changes pertaining to market microstructure and to predict instabilities such as flash crashes and short interest anomalies, and our work was covered by BoingBoing and the MIT technology review.

• Human-computer interaction: We patented our work on visualizing networked information, using which we developed an iOS app that achieved 100K+ downloads and was covered in Wired.

• Network science: We used nonlinear unsupervised learning methods to forecast instabilities in computer networks for the Department of Defense.

Talks and media

• Data Gurus Podcast, Feb 2020: Podcast on the importance of data quality.

• AI at Work Podcast, May 2019: Applying NLP to developing new SaaS products (see blog).

• MIT Sloan School Workshop, March 2019: Hands-on machine learning workshop.

• Deep Learning in Production Meetup, January 2019: Deep learning for NLP at HubSpot.

• Battle of the Bots, June 2018: Developing a differentiator in the bot race.

• BarCamp Boston, June 2011: Predictive analytics: the future of predicting the future.

• Times-Herald Record, June 2005: Brains + heart.

Writing

• ThinkGrowth, July 2017: AI and Big Data are changing our attention.

• Quora, November 2016: Understanding meditative ecstasy.

• Quora, November 2016: Cloud solutions for deep learning.

• Quora, December 2011: Understanding elevator scheduling algorithms.

• Quora, December 2011: Understanding Android exploits.

• Columbia Spectator, Feb 2009: How I learned to stop worrying and love the LHC.

• Columbia Spectator, March 2009: The Genomic atheneum.

• Columbia Spectator, Feb 2009: Bachelor of medicine.

• Columbia Spectator, Jan 2008: Giving a soft-spoken faith a louder voice.

Contact

Thanks for stopping by! You can find me at vedantmisra on Twitter or vedant on Github.