Preetum Nakkiran
I'm a 4th-year PhD student at Harvard, in the Theory Group and co-organizer of the Harvard ML Theory Group. I'm advised by Madhu Sudan and Boaz Barak. I recently interned at OpenAI, with Ilya Sutskever.
I'm currently working on advancing the science of deep learning, theoretically and empirically.
New:
- Our new paper Deep Double Descent, which demonstrates and unifies surprising phenomena in deep learning.
- My rebuttal piece "Adversarial Examples are Just Bugs, Too" on Distill.
- Our paper on the dynamics of SGD on neural-networks (NeurIPS 2019 spotlight).
- A short paper on adversarial examples in machine-learning.
Older:
- Our paper on using polar codes to efficiently compress Hidden-Markov sources, and code for related channels. (See slides from China Theory Week, and also our previous paper on polar codes).
- Our general-audience paper (and poster) on preventing false-discoveries in adaptive science.
- An exposition + demo of Gauss's Principle of Least Constraint.
I did my undergrad in EECS at UC Berkeley, and have previously interned with the speech-team at Google.
Contact:
preetum (at) cs.harvard.edu