Inverse Approximation of Non-linear Dimensionality Reduction

Informal Abstract

Dimensionality reduction (DR) is a technique in data analysis that compresses data to isolate the most important features and relationships. Linear DR refers to techniques which are reversible, such as PCA. Non-linear DR (or manifold learning) is not reversible, and creates a lossy embedding (i.e. data is lost). This project explores the use of interpolated radial basis functions (RBFs) to generically calculate approximate inverses of non-linear DR techniques.

Keywords: manifold learning, radial basis function, radial interpolant, particle swarm optimization