Browsing by Subject "Whitney embedding theorem"

Machine learning is by no means a novel field, but a recent boom in interest has led to a rapid increase in funding for related research. Because of this, many pure-mathematicians may find themselves trying to transition to this currently lucrative area of research. Thus, there is some demand for literature which helps ease this transition for mathematicians with a geometry or topology background. In this thesis we provide an introduction to contemporary machine learning research for a geometrically or topologically inclined individual. We do so by tracing the study of manifolds from their inception to modern machine learning. The thesis begins with a brief history of manifolds to motivate the examination of a proof of the Whitney embedding theorem. The theorem is then proved in detail, following texts from Adachi and Mukherjee on differential geometry. Later, a brief introduction to manifold learning introduces the reader to the manifold hypothesis and connects the classical study with its machine learning counterpart. Then, we provide a canonical introduction to neural networks after which we share rigorous mathematical definitions. Finally, we introduce the necessary preliminaries and subsequently prove the universal approximation theorem with injective neural networks. While we consider the Whitney embedding theorem as having applications in machine learning research, the universal approximation theorem with injective neural networks has clearer uses beyond mathematics. The studies of inverse problems and compressed sensing are two areas for which injectivity is a necessary condition for the well-posedness of common questions. Both fields have many deep applications to scientific and medical imaging. Injectivity is also a prerequisite for a function to preserve the topological properties of its domain.