Regularization by deep learning in signal processing

In this paper, we explore a new idea of using deep learning representations as a principle for regularization in inverse problems for digital signal processing. Specifically, we consider the standard variational formulation, where a composite function encodes a fidelity term that quantifies the proximity of the candidate solution to the observations (under a physical process), and a second regularization term that constrains the space of solutions according to some prior knowledge. In this work, we investigate deep learning representations as a means of fulfilling the role of this second (regularization) term. Several numerical examples are presented for signal restoration under different degradation processes, showing successful recovery under the proposed methodology. Moreover, one of these examples uses real data on energy usage by households in London from 2012 to 2014.