Harnessing Intrinsic Dynamics for Signal Processing

Nonlinear dynamics are a pervasive phenomenon in soft robotic system, where time-varying signals from different physics and stimuli in the environment influence the behavior of the system. Using the mathematical framework of reservoir computing, it is possible to harness these intrinsic, physical dynamics from the loading environment as a computational resource to augment their internal state assessment and control. Reservoir computing is a class of recurrent neural networks that trains only a readout layer of the network dynamics in contrast to tuning all the internal parameters of the network. This simplified training regime opens the door to physical implementations (vs on chip) of this neural network mapping behavior, and harnesses the unique, intrinsic nonlinearities and variabilities of a physical system. In this presentation, we will evaluate the information processing capacity of two distinct physical systems, specifically a optomechanical spring network and fluid flow in a cavity. We further develop a spectral projection method to characterize the relationship between system nonlinearity and reservoir computing performance. The analysis partitions the dimensionality increase of the signal among its frequencies, distinguishing between the strengths of the linear vs various nonlinear classes of frequency content. Together, this provides a framework to evaluate what types of output functions certain physics and material nonlinearities are best matched to process, and motivates the identification of new abstractions in more complex material systems, such as soft robotics, to directly perform embodied signal processing.