Characterizing and Improving the Robustness of Convolutional Neural Networks

Convolutional neural networks (CNNs) are expressive function approximators that play an important role in solving modern computer vision tasks, such as object recognition, and even summarizing images in natural language. Given their broad utility, CNNs have already been deployed in performance-critical systems, such as autonomous vehicles. Unfortunately, these models are vulnerable to subtle perturbations of their input, and typically have unreliable confidence estimates. These weaknesses have spawned a flurry of research aiming to devise reliable defense mechanisms, and tackle the confidence problem, but no compelling solution has been proposed to date. These open challenges severely limit the extent to which AI can be adopted in commercial settings that improve life and benefit the economy. This project has three goals: 1) characterize these limitations with respect to relevant concepts such as generalization and stability. TO BE CONT’D

Faculty Supervisor:

Graham Taylor

Student:

Angus Galloway

Partner:

Borealis AI

Discipline:

Engineering - other

Sector:

Information and communications technologies

University:

Program: