Game Theory-based Ensemble of Deep Neural Networks for Large Scale Audio Tagging

Audio tagging is concerned with the development of systems that are able to recognize sound events. With the growing interest geared towards audio tagging for various applications, it has become of paramount importance to design systems that distinguish among events of different natures. To mend with this, ensembling many tagging system has become a successful strategy that lives-up to these emerging challenges. In this paper, we introduce a tagging system composed of an ensemble of deep learners. We propose to formulate the fusion strategy as a coalitional game. Our approach weighs these individual learners, while considering two crucial notions that affect the performance of an ensemble: accuracy and diversity. To demonstrate the efficiency of our approach, we have carried out experimental comparisons on a huge dataset made of sound recordings with annotations of varying reliability. The experimental results indicate that the proposed system provides a reliable ranking and outperforms some major state-of-the art ensemble learning approaches.