Variational and PDE Approaches for Robust Artificial Intelligence (AI)

CS Talk

Variational and PDE Approaches for Robust Artificial Intelligence (AI)

Ganesh Sundaramoorthi
RTX Technology Research Center and Georgia Institute of Technology

Abstract:
Deep neural networks (DNNs) are ubiquitous in many real world applications.  Despite this success, in safety critical applications, there still exists barriers to adoption of this technology, because of a lack of robustness of these models and the theory of DNNs lags behind practice.  Towards building robust deep learning systems, we develop theoretical tools to study and develop robust DNNs.  We present recent work in two areas.  First, we present surprising numerical instabilities inherent in the training DNNs that lead to a lack of robustness, a theoretical analysis of this phenomena, and applications, e.g., to neural 3D scene representations from visual data.  Second, we examine robustness of DNNs online to changing data conditions through Out-of-Distribution Detection, where we present a theoretical framework that offers an explanation to existing methods, which are often times heuristically-driven.  The theory also leads to new methods that predict new methods that out-perform existing ones.   We introduce and develop tools based on variational and partial differential equations (PDE) methods to study and develop methods in these two areas.

Bio:
Ganesh Sundaramoorthi is a Senior Technical Fellow at RTX Technology Research Center (RTRC) and he is also an Adjunct Professor of Electrical & Computer Engineering at Georgia Institute of Technology.  His research is in fundamental and applied computer vision, machine learning, and artificial intelligence (AI), e.g., robustness, explainability, acceleration, and low size, weight & power.  He has played leading roles in several government programs, including the recent IARPA WRIVA program on 3D reconstruction and rendering from neural representations and the ARPA-E GOHPHURS program on 3D reconstruction from radar through neural representations.  He was area chair for leading AI conferences (IEEE/CVF CVPR & ICCV).  He has published more than 60 articles in leading AI venues and more than 30 patent applications.  His PhD was from Georgia Tech, and he was a postdoc at UCLA, both in computer vision / machine learning.  Prior to RTRC, he was on the faculty of KAUST, where he led a research group in computer vision.

Website: http://ganeshsun.com/