Studying Noise Sensitivity of Deep Neural Networks: An Analytic Approach

Bernard Ghanem
(KAUST, Saudi Arabia )

Abstract:

The first part of the talk will provide a brief overview of research done in the Image and Video Understanding Lab (IVUL) at KAUST, focusing on three primary themes:
  1. large-scale video understanding, e.g. fast/accurate human activity detection and object tracking,
  2. simulation for vision applications, e.g. self-driving cars and self-racing UAVs, and
  3. fundamentals of computer vision and machine learning, e.g. structured optimization (sparse, low-rank and binary), optimization inspired network design, and sensitivity analysis of deep networks.
The rest of the talk will delve into the details of recent work at IVUL on noise sensitivity analysis. This work provides an analytical framework to systematically measure the effect of Gaussian input noise on piecewise linear deep neural networks (PL-DNNs), e.g. networks with ReLU activations. Practically, this analysis can be exploited in several ways, including studying PL-DNN noise sensitivity, generating adversarial attacks, and training more robust PL-DNNs. Specifically, we derive exact and closed form analytic expressions for the 1st and 2nd moments (mean and variance) of a small PL-DNN (Affine->ReLU->Affine) subject to a general Gaussian input. We experimentally show that these expressions are tight under simple linearizations of deeper PL-DNNs, especially for well-known architectures (LeNet and AlexNet). Extensive experiments on image classification tasks show that these expressions can be used to directly study PL-DNN behavior under Gaussian input noise, including interclass confusion and pixel-level spatial noise sensitivity. Experiments also show that distributions (not mere samples) of successful adversarial Gaussian attacks (both targeted and non-targeted) can be computed by using the derived expressions to formulate and solve a suitably designed optimization problem.

Short Bio:

Bernard Ghanem is currently an Associate Professor of Electrical Engineering and Computer Science at King Abdullah University of Science and Technology (KAUST). He is also a theme leader at KAUST’s Visual Computing Center. Before that, he was a Senior Research Scientist at the University of Illinois Urbana-Champaign (UIUC) in Singapore. His research interests lie in computer vision, machine learning, and optimization geared towards real-world applications. He received his Bachelor’s degree in Computer and Communications Engineering from the American University of Beirut (AUB) in 2005 and his MS/PhD in Electrical and Computer Engineering from UIUC in 2010. His work has received several awards and honors, including two Best Workshop Paper Awards (CVPRW 2013 and ECCVW 2018), a two-year KAUST Seed Fund, and a Google Faculty Research Award in 2015. He has co-authored more than 70 peer reviewed conference and journal papers in his field

Visit ivul.kaust.edu.sa and http://www.bernardghanem.com for more details.