Physics Inspired Machine Learning
Speaker
Siddhartha Mishra, ETH Zurich
Time
2022-04-05 15:00 ~ 16:00, in 3 days (Asia/Shanghai Time)
Venue
Online
VooV Meeting Info
VooV Meeting download website:
https://voovmeeting.com/index.html Conference ID: 433633075
Password: 135448
ZOOM Info
Conference ID: 82675301781
Password: 123456
Abstract
Machine learning is being very successfully applied in the context of simulations of physical systems. We consider the reserve question: can one use physical principles and concepts to design robust and efficient machine learning systems? To this end, we will consider two situations. The first one deals with long range sequence modeling and we present recurrent sequential models that can learn tasks with long-term dependencies. Our architectures are based on physical systems (different flavor of oscillators and multi-scale ODEs). The second set of tasks we consider concern graph-structured data. In this context, we will present a graph neural network framework based on nonlinear coupled oscillators. In both cases, the use of physics inspired architectures solves fundamental problems in these settings (Exploding and Vanishing gradients for recurrent sequential models and over smoothing for Graph neural networks), while retaining sufficient expressivity to provide state of the art performance on a variety of learning tasks.
Bio
Siddhartha Mishra hails from Bhubaneswar, Odisha, India where he received his BSc in Physics and Mathematics from Utkal University in 2000. He obtained his MS degree in Mathematical Sciences (2003) and PhD in Mathematics (2005) from the Indian Institute of Science (IISc) and the Tata Institute of Fundamental Research (TIFR). Bangalore, India. He was a postdoctoral researcher from 2005 to 2009 at the Center of Mathematics for Applications (CMA) at the University of Oslo, Norway. From 2009 to 2011, Mishra was an assistant professor at SAM, ETH Zurich and an associate professor at the University of Oslo from 2011 to 2012. Since 2012, Mishra is at ETH Zurich, first as an associate professor from 2012 and as a Chair Professor of Applied Mathematics from 2015. Mishra is the Director of Computational Science Zurich and a core Faculty at the ETH AI Center.
Mishra's research interests are in numerical analysis, scientific computing, nonlinear PDEs, machine learning, computational fluid dynamics, computational astrophysics, computational climate science and modeling and simulation of biological systems. Mishra's research has been recognized with many awards and honors including the ERC Starting Grant (2012), ERC Consolidator Grant (2017), GAMM Richard Von Mises Prize (2015), ECCOMASS Jacques Louis Lions Medal (2018), ICIAM Collatz Prize (2019) and INFOSYS prize in Mathematical Science (2019). He has been an invited speaker at many leading international conferences including the International Congress of Mathematicians (ICM), Rio De Janiero, 2018.
Sponsors
Institute of Natural Sciences, Shanghai Jiao Tong University
Shanghai National Center for Applied Mathematics (SJTU Center)
Ministry of Education Key Lab in Scientific and Engineering Computing
--
FROM 211.161.249.*