Date of Award

5-2016

Document Type

Thesis

Degree Name

Master of Science (MS)

Legacy Department

Computer Engineering

Committee Member

Dr. Robert Schalkoff, Committee Chair

Committee Member

Dr. Harlan B. Russell

Committee Member

Dr. Yongqiang Wang

Abstract

Reducing the input dimensionality of large datasets for subsequent processing will allow the process to become less computationally complex and expensive. This thesis tests if Karnin sensitivity can be applied to reducing input dimensions of feed forward neural networks as well as comparing the results to the well known principal component analysis (PCA). The resulting error when reducing dimensions of inputs of various scenarios according to PCA and Karnin sensitivities are compared. After testing, Karnin was found to be able to be used to reduce input dimensions and did as well if not better than PCA in most cases. However, Karnin, like PCA, is not without its weaknesses. To cover both techniques' weaknesses, a combination of the two techniques is introduced. In the end, 'PCA chases variance while Karnin chases good mapping.'

Share

COinS