Seminars
View all Seminars | Download ICal for this eventImproving Global Robustness Checks by Approximating DNNs
Series: Department Seminar
Speaker: Kumar Madhukar
Date/Time: Apr 10 11:30:00
Location: CSA Seminar Hall (Room No. 254, First Floor)
Abstract:
Robustness is a crucial property for Deep Neural Networks (DNNs). It refers to a network??s
ability of ensuring reliable predictions even with small perturbations in the input. Unlike
local robustness, that checks this stability around a given input, global robustness provides
a comprehensive guarantee over the entire input space. Since it is a two-safety property, a
naive way to perform the global robustness check is through self-composition [Anagha et al., CAV 2024].
However, this doubles the size of the network and therefore does not scale very well. In this
talk, we look at a technique that suggests a marking for the neurons in the network based on
how they affect the final output(s) [Elboher et al., CAV 2020], and explore how such a marking can
help improve the scalability of global robustness checks.
Speaker Bio:
Kumar Madhukar is an Assistant Professor and Chandruka New Faculty Fellow in the Department of Computer Science and Engineering at IIT Delhi. His research interests are in Software Verification, Synthesis, Model Checking and Deep Learning.
Host Faculty: Deepak D'Souza