Visual analytics for improving deep learning multidimensional projections

Abstract

Deep learning has recently shown the ability to construct dimensionality reduction, or projections, with high quality and computational scalability. However, such methods have the major drawback of operating as black boxes, hence, it is hard for users to ?fine-tune them to achieve more specific projection styles. An important instance of this problem is the learning of t-SNE projections: The learned projections are typically fuzzier than the original t-SNE ones, making them less suitable for many visual analysis use-cases for which t-SNE was originally proposed. We aim to adapt and use classi?fier visualization methods to get a better understanding of the reasoning behind the network's inference of projections. We pinpoint, and apply ?fixes to ultimately reduce the causes of diffusion in the learned projections, culminating in the application of KNNP, a nearest-neighbors approach to the original NNP which further increases the quality of deep learning projections.