Learning of low-dimensional geometric structure in high-dimensional data

David Dunson (Duke University, USA)

High-dimensional data are very commonly collected in an amazing variety of application areas. As the sample size is often modest relative to the dimension of the data, it is necessary to develop dimensionality reduction methods. Most of these methods are based on assumptions of sparsity and/or linearity. For example, PCA reduces the original p-dimensional data to coordinates on a k-dimensional affine subspace, with k much less than p typically. There is a literature on non-linear extensions of PCA; such approaches are commonly referred to as "manifold learning" - supposing that the lower-dimensional subspace consists of one or more smooth Riemannian manifolds. Almost all such approaches rely on the assumption of local linearity, which can be justified based on the manifolds being approximately Euclidean in local neighborhoods. The key problem with such approaches is that a very large number of linear pieces are needed if the manifold(s) have moderate to high curvature; for example, think of approximating a low-dimensional sphere with planes. To address this problem, we propose a broad new dictionary for approximating manifolds based on pieces of spheres or spherelets. We provide theory showing dramatic reductions in covering numbers needed to produce a particular small MSE relative to locally linear methods. We develop a simple spherical PCA algorithm for implementation, and show very substantial gains in performance relative to the state-of-the-art on toy and non-toy examples. We additionally develop a Bayesian model that characterizes uncertainty in subspace learning, along with a simple and efficient MCMC algorithm for implementation.