Neural networks for inverse control with system priors

Anonymous

Keywords: imaging, inverse problem, deep learning, untrained networks, information retrieval, phase retrieval, scattering media

TL;DR: An algorithm for obtaining the required input of an arbitrary system that produces the desired output without labeled data with untrained networks

Abstract: Controlling a physical system to behave in a desired manner requires a prior knowledge of the system, having access to a model or a collection of labeled data consisting of the desired inputs-outputs of the system. When the system is unknown or labeled data is not available or expensive to acquire, resorting to approaches that do not rely on the use of training data is inevitable. In this work, we propose an algorithm based on untrained neural networks that can be applied to a physical system in its most general form to obtain the required input that would result in a desired target output. We showcase the applicability of our algorithm to experimental phase-retrieval problems in the complex environment of a scattering medium whose input-output relation follows a nonlinear and slowly time-varying setting. We show that despite partial measurements of the system, comparable fidelity to that of fully-observed methods or supervised networks is achievable.