Faculty Advisor

Paffenroth, Randy Clinton

Abstract

In this MQP project, our focus is on inverse methods for non-linear manifold learning. Methods for taking high-dimensional data and non-linearly projecting it to low-dimensions are well known. However, methods for going the other direction are less well studied. Here, we examine the use of linear and nonlinear methods of dimension reduction and study the particulars of the inverse mapping. Furthermore, we study which combinations of parameters are more effective when there are restrictions on our projections' dimension. We explore a variety of different projection algorithms including Principal Component Analysis, Isomap, and Local Tangent Space Alignment. Finally, the effectiveness of our approaches will be demonstrated on the image recognition problem of classifying handwritten digits.

Publisher

Worcester Polytechnic Institute

Date Accepted

May 2016

Major

Mathematical Sciences

Project Type

Major Qualifying Project

Accessibility

Unrestricted

Advisor Department

Mathematical Sciences

Share

COinS