Adaption of Layerwise Relevance Propagation for Audio Applications (vor Ort)
* Presenting author
Deep neural networks (DNNs) excel at signal detection and parameter estimation tasks, particularly under conditions where traditional algorithms struggle. In contrast to analytically derived methods, DNNs solve tasks through their parameters being optimized for a particular task in training. However, this comes at the cost of a DNN model's operation being intransparent – behaving like a black box.Layerwise Relevance Propagation (LRP) is a method originally developed in the field of image recognition which aims to resolve this black box behavior by attributing relevance to a model’s input features, representing the contributions of these features to the model’s prediction for a particular input. Analysis of the resulting heatmaps is then to give insight regarding the model’s decision strategy.This work examines the specifics of applying LRP to DNNs used for audio applications – from a view on specific properties of audio signals as compared to images, through ways of evaluating LRP heatmaps qualitatively and quantitatively by means of visualization, auralization and evaluation of multiple metrics. The interpretation of LRP results and their utility in DNN model design is discussed. LRP is applied exemplarily to DNNs for Direction of Arrival (DOA) estimation from Higher Order Ambisonics signals.