Physical learning machines use physical dynamics to achieve the same kind of general information processing and training that artificial neural networks are known for, but possibly faster and more energy-efficient. Self-learning machines can be defined as an ambitious subset of physical learning machines: They can be trained without requiring any form of external feedback to update the trainable parameters inside the device. In this talk, I will describe a new general idea that can be used to realize self-learning starting from any kind of Hamiltonian system with time-reversible dynamics. I will explain the reasoning and intuitive ideas behind it, as well as the ingredients that will be useful in building such devices in various physical platforms.

Since 2016, Florian Marquardt leads the theory division of the Max Planck Institute for the Science of Light in Erlangen, Germany. His work covers the intersection of nanophysics and quantum optics. Research topics include cavity optomechanics, topological transport, quantum many-body dynamics, and the interface between machine learning and physics. After defending his thesis in 2002 in Basel, Switzerland, he was a postdoctoral fellow at Yale before becoming a junior research group leader at the Ludwig-Maximilians-Universität Munich and finally a full professor and subsequently Max Planck director at Erlangen.