Manifold fitting under unbounded noise
This interesting paper deals with a classical problem in machine learning: fitting a manifold to data. Suppose we are in Euclidean space \(\mathbb R^n\) with high dimension \(n\) and a collection of data points are drawn from an unknown probability distribution in \(\mathbb R^n\). The manifold hypothesis asks when it is true that this data lives on or is close to an intrinsically low-dimensional embedded manifold. A classical example of when the manifold hypothesis holds occurs in computer vision where a picture of a face, say, consists of millions of pixels (the dimension of the ambient Euclidean space). Other examples of the manifold hypothesis holding occur in medical imaging, seismology and many others. This paper studies manifold fitting when the data is corrupted by noise. The ultimate aim of manifold fitting is to represent the latent manifold as an embedded sub-manifold of the ambient space. Once the latent manifold is learned, various types of analyses can be carried out based on it, such as denoising and many others. Given various assumptions on the noise (such as concentration and boundedness), a common method to construct the manifold is based on a tangent space estimation at each sample point, which gives good theoretical results if, apart from other conditions, the noise is bounded. If the noise is unbounded, the tangent estimation might well be blurred and so the manifold fitting method is compromised. \N\NThis paper introduces a manifold fitting method where, in this case, instead of the output manifold being constructed based on a tangent space estimation at each sample point, the output manifold is constructed using certain tangent spaces and projections of points onto the latent manifold itself, thus reducing the error caused by the noise. If the noise is unbounded the method has a high probability of achieving theoretical convergence, in terms of the upper bound of the distance between the estimated and latent manifold.\N\NNumerous numerical experiments validate the papers new theoretical ideas and findings.\N\NThe paper is well written with a good set of references.
- A k-segments algorithm for finding principal curves
- Curvature Measures
- Finding the homology of submanifolds with high confidence from random samples
- Locally defined principal curves and surfaces
- Manifold estimation and singular deconvolution under Hausdorff loss
- Manifold reconstruction using tangential Delaunay complexes
- Minimax manifold estimation
- Nonasymptotic rates for manifold, tangent space and curvature estimation
- Nonparametric ridge estimation
- On Eigenvalue Optimization
- Principal boundary on Riemannian manifolds
- Principal Curves
- Random Fixed Boundary Flows
- Stability and minimax optimality of tangential Delaunay complexes for manifold reconstruction
- Testing the manifold hypothesis
- The geometry of nonparametric filament estimation
- The reach, metric distortion, geodesic convexity and the variation of tangent spaces
This page was built for publication: Manifold fitting under unbounded noise
Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q6982657)
