Author

Wiley Corning

Date of Award

2017

Document Type

Thesis

Degree Name

Bachelors

Department

Natural Sciences

First Advisor

McDonald, Patrick

Area of Concentration

Mathematics

Abstract

Although neural networks are currently used in many applications and research environments, they remain poorly understood as mathematical objects. In this thesis, we investigate the topological and algebraic properties of neural networks. We develop an understanding of algebraic structure of neural networks and produce a novel distance metric on the parameter space. We derive a backpropagation algorithm to compute the Hessian matrix of a deep rectifier network. We perform a synthetic data experiment to explore the error landscape of simple networks, and, using our distance metric, find clear patterns in the spatial distributions of learned parameters.

Share

COinS