Alhussein Fawzi Seminar at the Department of Engineering Science, University of Oxford

Image classification systems have recently witnessed huge accuracy gains on several complex benchmarks. However, while being very accurate, these classifiers are not robust to slightest perturbations of the data. In this talk, I will first highlight the vulnerability of state of the art classifiers in a broad set of perturbation regimes, such as adversarial and universal perturbations.

After presenting ways to improve the robustness of these classifiers, I will describe the existence of a close relation between the robustness of classifiers and their geometry in the input space. This connection can be leveraged to further improve the robustness to adversarial perturbations. I will conclude with open problems in the field.

Bio: 

Alhussein Fawzi is a senior research scientist at DeepMind in London working on the robustness of machine learning, and more recently on solving combinatorial problems using machine learning. He received his M.Sc. and PhD degrees from the Swiss Federal Institute of Technology (EPFL), Switzerland, and spent one year as a postdoctoral scholar in the Computer Science Department at UCLA. He received twice the IBM PhD fellowship. More information can be found in his website: http://www.alhusseinfawzi.info/