SEMI-SUPERVISED SOURCE LOCALIZATION WITH RESIDUAL PHYSICAL LEARNING

Machine learning (ML) approaches to source localization have demonstrated promising results in addressing reverberation. Even with large data volumes, the number of labels available for supervised learning in such environments is usually small. This challenge has recently been addressed using semi-supervised learning (SSL) based on deep generative modeling with variational autoencoders. A problem with ML approaches is they often ignore the intuitions from conventional signal processing approaches. We present a hybrid approach to ML-based source localization, which uses both SSL and conventional, analytic signal processing approaches to obtain source location estimates. An SSL approach is developed which accounts for the residual between analytic source location estimates true locations. Thus, the approach can exploit both labelled and unlabeled data, as well as analytic source location intuition, to provide better localization than either approach in isolation.

Automated summary

Machine learning approaches have been successful in source localization, but often suffer from limited labeled data. By combining semi-supervised learning and traditional signal processing, a hybrid approach can achieve better source localization.