Event:

Artificial intelligence is transforming science and society, yet current systems are far from reliable: they fail in unexpected ways, their decisions are hard to explain, and their energy demands are growing at an unsustainable pace. In this talk I will argue that a mathematical theory of AI is essential to address these challenges. I will first give an overview of what we currently understand about deep neural networks – their expressivity, their training, their generalization, and their explainability – and show how these insights translate into practical guidelines for building trustworthy systems. I will then turn to a fundamental limitation: on today’s digital hardware, several key problems in AI turn out to be provably non-computable, placing hard limits on the reliability that can be achieved. This points toward analog and neuromorphic hardware, and in particular toward spiking neural networks, which are far closer to biological computation. I will present recent results on the expressivity and computational power of spiking networks and discuss what a mathematical theory of brain-inspired computing could look like. The talk is intended for a broad audience and will not assume a background in mathematics.