Seminars

View all Seminars  |  Download ICal for this event

What does PAC-learning have to say about adversarial robustness?

Series: Bangalore Theory Seminars

Speaker: Vinayak Pathak

Date/Time: Aug 19 17:00:00

Location: CSA auditorium [Room No. 104], ground floor

Abstract:
Modern neural network architectures have achieved great success on various learning tasks, but they are almost always prone to adversarial attacks. These attacks are small, undetectable perturbations of the input that lead the model to change its output. The widespread existence of adversarial attacks is both a security concern, as well as an indication that these models arent learning the ground truth (since the true label does not change when input is perturbed by a small amount).

PAC-learning is a theoretical framework that has been used to study learning problems for decades. Several recent papers have analyzed under this framework the question of learning in a way thats immune to adversarial attacks. This investigation has led to several interesting results; however, many practical phenomena still remain unexplained. In this talk I will provide a brief introduction to PAC-learning and discuss a few approaches to formulating the problem of adversarially robust learning under this framework. I will survey some recent papers and discuss open questions.


Microsoft teams link:
Link

We are grateful to the Kirani family for generously supporting the theory seminar series


Hosts: Rameesh Paul, KVN Sreenivas, Rahul Madhavan, Debajyoti Kar