PIMS - ULethbridge Distinguished Speaker Series: Ben Adcock

Deep learning lies at the forefront of the artificial intelligence revolution. Stunning successes has been achieved by deep learning for challenging tasks such as image classification. Yet, current deep learning implementations have a tendency to be unstable, and vulnerable to so-called adversarial attacks. In the first part of this talk, I will give an overview of these instabilities, and their potential consequences in applications. In the second part, I aim to focus on a rather different application of deep learning: namely, inverse problems in imaging. Image reconstruction is a crucial task in many different scientific, industrial and medical technologies, and one to which deep learning techniques have recently begun to be applied. However, as I will demonstrate, instabilities persist in this problem as well. I explain how such instabilities can be constructed, and conclude with some theoretical insights into why they arise and how they may eventually be mitigated.

This talk requires no prior knowledge of neural networks and should be accessible a general audience at the graduate or senior undergraduate level.