Deep Unfolding for Data Estimation in Wireless Communication Systems
Sprache des Vortragstitels:
Englisch
Original Tagungtitel:
SIAM Conference on Computational Science and Engineering 2023 (CSE23)
Sprache des Tagungstitel:
Englisch
Original Kurzfassung:
Traditionally, data estimation on the receiver side of a wireless digital communication system is accomplished with model-based methods. These methods are based on well-established physical and statistical models. Consequently, they are well-interpretable and performance bounds can often be derived. However, modeling errors, oversimplifications, wrong statistical assumptions, or insufficient model knowledge may severely degrade the performance of model-based approaches, and incorporating empirical statistics of possibly available data is usually difficult. Data-driven approaches can resolve some of the aforementioned issues. However, they usually suffer from huge data hunger, and they typically lack interpretability. Some of these problems may be tackled by incorporating model knowledge into data-driven methods, which is a major challenge with lots of open research questions.
In this talk, we present model-inspired neural networks (NNs) for data estimation, which are derived by using deep unfolding. These NNs are designed by unfolding the iterations of iterative model-based algorithms to layers of NNs. We highlight similarities between model-based methods and model-inspired NNs. For example, we show that conducting preconditioning, which is known to improve iterative model-based methods, can boost the performance of NNs that are derived by deep unfolding. We compare these NNs to traditional model-based methods, and highlight their pros and cons.