Espadoto, MateusRodrigues, Francisco Caio MaiaHirata, Nina S. T.Hirata Jr., RobertoTelea, Alexandru C.Landesberger, Tatiana von and Turkay, Cagatay2019-06-022019-06-022019978-3-03868-087-1https://doi.org/10.2312/eurova.20191118https://diglib.eg.org:443/handle/10.2312/eurova20191118We present a new method for computing inverse projections from 2D spaces to arbitrary high-dimensional spaces. Given any projection technique, we train a deep neural network to learn a low-to-high dimensional mapping based on a projected training set, and next use this mapping to infer the mapping on arbitrary points. We compare our method with two recent inverse projection techniques on three datasets, and show that our method has similar or higher accuracy, is one to two orders of magnitude faster, and delivers result that match well known ground-truth information about the respective high-dimensional data. Visual analytics Unsupervised learning Dimensionality reduction and manifold learning.VisualizationVisualization application domainsMachine learningLearning paradigms10.2312/eurova.2019111813-17