In deep learning, artificial neural networks are trained to perform mathematical functions solely by providing labelled examples of the desired transformation, and then incrementally updating the neural networks' parameters example by example so the network "learns" to performs the function better and better. In my talk, I will describe how one can treat more or less any physical system analogously, training it to perform desired physical functions by machine-learning its physical parameters. For example, we have trained physical systems to perform analog computations, such as a metal plate whose modulated oscillations can classify images. We have trained physical systems to act as sensors, which can extract desired information from the physical world - for example, we have trained an optical system to sense numbers from a scene in front of it. Of course, this can also be applied to many other types of physical functions: one can similarly machine-learn physical tools that perform desired operations, or even entire physical experiments.
Host: Naomi Gluck
