Pratik Chaudhari

I recently defended my PhD thesis in the Computer Science department at UCLA and I worked with Stefano Soatto in the Vision Lab.

I will be joining the faculty in the ESE department and the GRASP Lab at the University of Pennsylvania starting July 2019.

I have Engineer's and Master's degrees in Aeronautics & Astronautics from MIT where I worked with Emilio Frazzoli at the Laboratory of Information and Decision Systems (LIDS). I was in the Aerospace Engineering department at IIT Bombay for my undergraduate studies until 2010.

I am interested in deep learning, robotics and computer vision. I draw from physics, optimization and probability to solve problems in these domains. I have worked extensively on self-driving cars in the areas of computer vision, motion-planning and stochastic estimation at nuTonomy Inc.

Resume CV

Contact

  pratik dot ac at gmail dot com

Web

GitHub Twitter Facebook Google+

Publications (Google Scholar)

	Stochastic gradient descent performs variational inference, converges to limit cycles for deep networks P. Chaudhari, S. Soatto arXiv:1710.11029 International Conference of Learning and Representations, 2018 Video of a talk at IPAM New Deep Learning Techniques workshop Short version PDF at Advances in Approximate Bayesian Inference, NIPS 2017 SIAM Conference on Imaging Science, 2018
	Parle: parallelizing stochastic gradient descent P. Chaudhari, C. Baldassi, R. Zecchina, S. Soatto, A. Talwalkar, A. Oberman arXiv:1707.00424 Code Short version at SysML Conference, Stanford, 2018
	Deep Relaxation: partial differential equations for optimizing deep neural networks P. Chaudhari, A. Oberman, S. Osher, S. Soatto, G. Carlier arXiv:1704.04932 Research in the Mathematical Sciences (RMS) Short version PDF at Principled Approaches to Deep Learning, ICML 2017 SIAM Conference on Analysis of Partial Differential Equations, 2017 Asilomar Conference on Signals, Systems, and Computers, 2017
	Entropy-SGD: Biasing gradient descent into wide valleys P. Chaudhari, A. Choromanska, S. Soatto, Y. LeCun, C. Baldassi, C. Borgs, J. Chayes, L. Sagun, R. Zecchina arXiv:1611.01838 International Conference of Learning and Representations, 2017 Code
	On the energy landscape of deep networks P. Chaudhari, S. Soatto arXiv:1511.06485 Advances in non-convex analysis and optimization, ICML 2016
	Incremental synthesis of minimum-violation control strategies for robots interacting with external agents P. Chaudhari, T. Wongpiromsarn, E. Frazzoli PDF American Control Conference (ACC), 2014 Code
	Sampling-based algorithms for optimal motion planning using process algebra specifications P. Chaudhari, V. Varricchio, E. Frazzoli PDF IEEE Conference on Robotics and Automation (ICRA), 2014 Video
	Game theoretic controller synthesis for multi-robot motion planning Part I : Trajectory based algorithms M. Zhu, M. Otte, P. Chaudhari, E. Frazzoli arXiv:1402.2708 IEEE Conference on Robotics and Automation (ICRA), 2014
	Incremental sampling-based algorithm for minimum-violation motion planning L. Reyes-Castro, P. Chaudhari, J. Tumova, S. Karaman, E. Frazzoli, D. Rus arXiv:1305.1102 IEEE Conference on Decision and Control (CDC), 2013 Code Watch a video of the demonstration here. Most societally beneficial video at International Joint Conference on Artificial Intelligence (IJCAI), 2014.
	Sampling-based algorithms for continuous-time POMDPs P. Chaudhari, S. Karaman, D. Hsu, E. Frazzoli PDF American Control Conference (ACC), 2013 Code
	Sampling-based algorithm for filtering using Markov chain approximations P. Chaudhari, S. Karaman, E. Frazzoli PDF IEEE Conference on Decision and Control (CDC), 2012 Code

Theses

	Algorithms for autonomous urban navigation with formal specifications P. Chaudhari PDF Engineers's thesis, Aeronautics and Astronautics, MIT, 2014
	Incremental sampling based algorithms for state estimation P. Chaudhari PDF Master's thesis, Aeronautics and Astronautics, MIT, 2012

Talks/Posters

• Unraveling the mysteries of stochastic gradient descent on deep networks
  New Deep Learning Techniques (IPAM, UCLA), Information theory and Applications (ITA18)
• A picture of the energy landscape of deep neural networks
  Stanford, MIT, Scholass Dagstuhl (Germany), Amazon AWS, OpenAI, CDC 2017
• Entropy-SGD: Biasing gradient descent into wide valleys
  IPAM, Statistics (UCLA), ENS Paris and INRIA, ENS Cachan, Politecnico di Torino, Information theory and applications (ITA17)
• Spin glasses: on the energy landscape of deep networks
  NYU, UCLA CS 269 ML seminar
• Visual representations: Defining properties and Deep approximations
  ICLR 2016
• Sampling-based algorithms: stochastic systems and formal specifications
  KTH, RSS, Qualcomm Research, Nissan
• Approximate POMDP homomorphisms
  MIT