MPI-DS Colloquium: Physical machines that can be made to learn

The exponential explosion of resources now being invested in machine learning and artificial intelligence are on an unsustainable trajectory. That is why a community of researchers is trying to design and explore physical devices that are able to mimic the functionality of digital artificial neural networks, but with greatly reduced energy needs and with natural parallel processing of data. One central question in this setting is how the physical dynamics and interactions can be used to implement efficient physics-based training. In this talk, I will give several examples of ideas we have introduced in this area over the past few years and describe the outlook for the future.