FREQUENCY DOMAIN BEAMFORMING USING NEURONAL NETWORKS

Abstract

In our contribution, we examine the use of Neuronal Networks for beamforming in the frequency domain. While conventional beamformers and deconvolution algorithms are often a compromise of spatial resolution, assumptions about the Green’s function, and computational effort, machine learning offers a new perspective on beamforming. While there exists some research that shows the potential of Neuronal Networks for beamforming, this study focuses on the different use-cases, and challenges that arise from the corresponding architectures. We discuss possible input and output designs for Neuronal Networks, different layer designs, and different activation-, and loss functions. The different architectures are evaluated and discussed for several source types such as monopoles and dipoles using several metrics for the quality of the resulting maps.