Model-Based Deep Learning

Signal processing, communications, and control have traditionally relied on classical statistical modeling techniques. Such model-based methods utilize mathematical formulations that represent the underlying physics, prior information, additional domain knowledge. Simple models are useful but sensitive to inaccuracies may lead poor performance when real systems display complex or dynamic behavior. On other hand, purely data-driven approaches model-agnostic becoming increasingly popular as datasets become abundant power of modern deep learning pipelines increases. Deep neural networks (DNNs) use generic architectures learn operate from data demonstrate excellent performance, especially for supervised problems. However, DNNs typically require massive amounts immense computational resources, limiting their applicability some scenarios. In this article, we present leading studying designing systems. These combine principled with benefit advantages both approaches. exploit partial knowledge, via structures designed specific problems, limited data. Among applications detailed in our examples compressed sensing, digital tracking state-space models. Our aim is facilitate design study future at intersection signal processing machine incorporate domains.