All Times ET
Keywords: Manifold learning, Dimensional reduction, Visualization, Deep learning
In this paper, we develop an inductive manifold learning algorithm that is ideal for dimensional reduction and visualization with deep neural net- works (DNNs), which generalizes on the unseen test dataset. Our method preserves both global and local information by approximating geodesic distances on the data manifold and preserving them. Theoretical justification on the distance estimator is provided with a careful analysis of the topological structure of the manifold.