Emergent learning in physical systems as feedback-based aging in a glassy landscape

Given the prevalence of emergent behavior, physicists, computer scientists, and biologists have long asked whether or not some subset of emergent behavior results in the capacity of a system of many interacting components to learn, i.e., to have intelligence [1, 2]. While there has been much focus looking for emergent learning in brain-like systems, such as neuronal networks in biology or artificial neural networks in physics and computer science, recent research has demonstrated that simple physical systems, such as a spring network, have the potential to exhibit learning behavior similar to that of artificial neural networks [3-9]. In this context, learning refers to the ability to modify the properties of a physical system by adjusting its learning degrees of freedom in order to more efficiently achieve some task. For example, in a spring network, the spring stiffness and rest lengths represent the learning degrees of freedom, while the nodes of the springs correspond to the usual physical degrees of freedom. In these physical learning systems, once input boundary nodes, output boundary nodes, and a cost function are all chosen, the learning process is composed of two steps: 1. Signaling: System's response to a given input is compared with the desired output and an update signal is sent which provides information on the necessary adjustments to each learning degree of freedom, so that the system's response aligns more closely with the desired output.