Topic of Research Seminar: Learning Neural Network Under General Specifications

Abstract: This work focuses on examining an important problem of learning neural networks that certifiably meet a certain class of convex relaxable specifications. The strategy is to find an inner approximation of the set of admissible policy parameters, which is convex in a transformed space. To this end, we address the key technical challenge of convexifying the verification condition for neural networks, which is derived by abstracting the nonlinear specifications and activation functions with quadratic constraints resulting in semidefinite programming (SDP). In particular, we propose a reparametrization scheme of the original neural network based on loop transformation, which leads to a convex condition that can be enforced during learning. We significantly push the scalability frontier of an SDP while achieving zero accuracy loss by leveraging chordal sparsity– decomposing large semidefinite constraints into equivalent smaller ones, and a tunable parameterization allowing us to trade off efficiency and accuracy in the learning process.

Subject Field of Topic: Machine Learning

Name of Speaker: Zain ul Abdeen

Professorial Rank of Speaker: PhD Student

University Email of Speaker: [email protected]

Name of the School organizing the Seminar: NUST School of Natural Sciences (NUST-SNS)

Speaker Profile weblink: https://zainulabdeen6.wordpress.com/

Affiliation of Speaker: Virginia Tech, USA

Date and Venue: March 15, 2023, at 1530 hrs, in CR 303, School of Natural Sciences (SNS), NUST Islamabad Campus