Date: Thursday, October 26, 2023 

Time: 12:00 PM – 1:30 PM EST 

Virtual Link: Zoom Link 

Meeting ID: 943 0118 3556 

Passcode: 766695 

Junjiao Tian 

Robotics Ph.D. Student 

School of Electrical and Computer Engineering 

Georgia Institute of Technology  

Committee

Dr. Zsolt Kira (Advisor) – Interactive Computing/Robotics, Georgia Institute of Technology 

Dr. Judy Hoffman – Interactive Computing/Robotics, Georgia Institute of Technology 

Dr. Animesh Garg – Interactive Computing/Robotics, Georgia Institute of Technology 

Dr. Jean Oh – Robotics Institute, Carnegie Mellon University  

Dr. Dustin Tran – DeepMind, Google  

Abstract

Sensitivity and robustness are often two sides of the same coin. They refer to different expected behaviors in the output of a system, responding to changes in the input. Specifically for a deep learning vision system, using classification as an example, a sensitive system should reflect the confidence (uncertainty) of its decision caused by the variations while a robust system should recognize an object under different variations. Practically, sensitivity and robustness both contribute to the safety of the deployment of deep learning systems in the real world. However, they do not naturally emerge in a conventional training pipeline. On the contrary, naive training methods often encourage the opposite. In this thesis proposal, I study the causes of losing sensitivity and robustness during training and propose strategies to improve them for safer applications of deep vision systems.