Bringing the Discussion of Minima Sharpness to the Audio Domain: a Filter-Normalised Evaluation for Acoustic Scene Classification

Intuitively, these terms are related in the context of deep neural networks has been subject to the Hessian matrix, which contains all second-order derivatives, at to discussion for a long time. Whilst mostly investigated in the a given point of a function, for all directions and can thus represent context of selected benchmark data sets in the area of computer vision, the local curvature behaviour of the function. Yet, an undisputed definition we explore this aspect for the acoustic scene classification task of flatness and sharpness in the high-dimensional parameter of the DCASE2020 challenge data. Our analysis is based on twodimensional space of ANNs is still lacking. Nevertheless, several approaches to filter-normalised visualisations and a derived sharpness quantify flatness and sharpness have been developed over the years, measure. Our exploratory analysis shows that sharper minima tend but they have failed to paint a complete picture of the generalisation to show better generalisation than flat minima -even more so for capabilities based on geometry, as a universal correlation between out-of-domain data, recorded from previously unseen devices-, thus flatness and generalisation has been disputed [6, 7].