Public Lecture – Deep Learning: What Could Go Wrong?

A traffic “Stop” sign on the roadside can be misinterpreted by a driverless vehicle as a speed limit sign when minimal graffiti is added. Wearing a pair of adversarial spectacles can fool facial recognition software into thinking that we are Brad Pitt. The vulnerability of artificial intelligence (AI) systems to such adversarial interventions raises questions around security and ethics, and many governments are now considering proposals for their regulation. I believe that mathematicians can contribute to this landscape. We can certainly get involved in the conflict escalation issue, where new defence strategies are needed to counter an increasingly sophisticated range of attacks. Perhaps more importantly, we also have the tools to address big picture questions, such as: What is the trade-off between robustness and accuracy? Can any AI system be fooled? Do proposed regulations make sense? Focussing on deep learning algorithms, I will describe how mathematical concepts can help us to understand and, where possible, ameliorate current limitations in AI technology.