Pratik Rathore

I am a fourth-year PhD student in the Electrical Engineering department at Stanford University, interested in optimization for machine learning, advised by Madeleine Udell.

Before Stanford, I graduated from the University of Maryland with a double degree in Electrical Engineering and Mathematics.

I most recently interned at Gridmatic, where I worked on faster optimization for battery scheduling and price impact models for energy trading.

Email  /  CV  /  Google Scholar  /  LinkedIn  /  GitHub

I'm interested in using randomization and preconditioning to design fast, scalable optimization algorithms for machine learning. Recently, I've been thinking about optimization challenges in scientific machine learning and using randomized low-rank approximations to develop new algorithms for convex, finite-sum optimization and deep learning.

^* denotes equal contribution.

We develop ASkotch, a scalable, accelerated, iterative method for full kernel ridge regression (KRR) that provably obtains linear convergence. ASkotch outperforms state-of-the-art KRR solvers on a testbed of 23 large-scale KRR regression and classification tasks derived from a wide range of application domains. Our work opens up the possibility of as-yet-unimagined applications of full KRR across a number of disciplines.

We study challenges in training physics-informed neural networks. We link training issues to ill-conditioning of the loss, and show a combined Adam and L-BFGS approach, along with a new optimizer, NysNewton-CG, enhances PINN performance.

We propose PROMISE, a family of preconditioned stochastic optimization methods that use scalable, randomized curvature estimates to solve large-scale, ill-conditioned convex optimization problems in machine learning. PROMISE methods, with default hyperparameters, outperform popular tuned stochastic optimizers on ridge and logistic regression. Furthermore, we introduce quadratic regularity, which determines the speed of linear convergence for PROMISE methods and allows us to obtain improved rates for ridge regresison.

We use techniques from randomized numerical linear algebra to develop a fast, scalable, preconditioned stochastic gradient method for convex machine learning problems.

We prove new results regarding the prime factorizations of Descartes numbers, a family of odd spoof perfect numbers.

Template borrowed from Jon Barron.