Latent Space Cartography for Geometrically Enriched Latent Spaces

Abstract

AbstractThere have been many developments in recent years on the exploitation of non-Euclidean geometry for the better representation of the relation between subgroups in datasets. Great progress has been made in this field of Disentangled Representation Learning, in leveraging information geometry divergence, manifold regularisation and geodesics to allow complex dynamics to be captured in the latent space of the representations produced. However, interpreting the high-dimensional latent spaces of the modern deep learning-based models involved is non-trivial. Therefore, in this paper, we investigate how techniques in Latent Space Cartography can be used to display abstract and representational 2D visualisations of manifolds.Additionally, we present a multi-task metric learning model to capture in its output representations as many metrics as is available in a multi-faceted fine-grained change detection dataset. We also implement an interactive visualisation tool that utilises cartographic techniques that allow dimensions and annotations of graphs to be representative of the underlying factors affecting individual scenarios the user can morph and transform to focus on an individual/sub-group to see how they are performing with respect to said metrics.