Inverse Design for Learning in Soft Matter Systems

Training an artificial neural network requires solving an inverse problem, so it is natural to apply concepts underlying machine learning to solve inverse design problems in soft matter systems. I will describe how global minimization of a loss function can be used to teach physical networks how to perform functions inspired by biology, such as the ability of proteins (e.g. hemoglobin) to change their conformations upon binding of an atom (oxygen) or molecule, or the ability of the brain’s vascular network to send enhanced blood flow and oxygen to specific areas of the brain associated with a given task. Statistical physics teaches us that the ability to design ensembles of networks with the same function is crucial in order to understand the microscopic origins of collective phenomena. I will show how we apply persistent homology to such networks to learn how and what they learned. Finally, I will describe local learning methods that allow us to develop physical systems that can learn on their own, without requiring a processor or memory.