Tuesday, 26th of March

14h30 (room R2014, 660 building) (see location)

Saumya Jetley

(University of Oxford)

Title: DeepInsight - An examination of the class decision functions learned by deep nets


Abstract

The talk will uncover some simplistic facets of the class score functions learned by deep image classification networks that explain their adversarial vulnerability and tie it to the celebrated performance of these nets. More particularly, our research identifies the fact that specific input image space directions tend to be associated with fixed class identities. This means that simply increasing the magnitude of correlation between any input image and a single image space direction causes the net to believe that more (or less) of the class is present. This provides a new perspective on the existence of universal adversarial perturbations for deep nets. Further, in the space mapped by the above directions, the adversarial vulnerability of the nets and their classification accuracy are closely entwined. Various notable and interesting observations emerge from this, with key implications for efforts to construct neural nets that are both accurate and robust to adversarial attack, and will be discussed in a greater detail during the talk.

Bio

Saumya Jetley is a final year PhD student under Prof. Philip Torr at the University of Oxford. Her research interests include machine vision, deep learning, attention models and more recently the geometrical analysis of decision boundaries of deep image classification networks to understand their adversarial vulnerability. She joined the University of Oxford as a Masters (in research) student in 2014 and stayed on for her PhD. Previously, she has had 3 years of industrial research experience from working with a Govt. of India research organisation (C-DAC) and was a summer intern at Xerox Research Centre Europe in 2015.



Contact: guillaume.charpiat at inria.fr
All TAU seminars: here