ORCID
https://orcid.org/0000-0002-9319-5630
Date of Award
Fall 1-10-2020
Degree Name
Master of Science (MS)
Degree Type
Thesis
Abstract
Artificial neural networks(ANNs) are recognized as high-performance models for classification problems. They have proved to be efficient tools for many of today's applications like automatic driving, image and video recognition and restoration, big-data analysis. However, high performance deep neural networks have millions of parameters, and the iterative training procedure thus involves a very high computational cost. This research attempts to study the relationships between parameters in convolutional neural networks(CNNs). I assume there exists a certain relation between adjacent convolutional layers and proposed a machine learning model(MLM) that can be trained to represent this relation. The MLM's generalization ability is evaluated by the model it creates based only on the knowledge of the initial layer. Experiments and results show that the MLM is able to generate a CNN that has very similar performance but different in parameters. In addition, taking advantage of the difference, I insert noise when creating CNNs from the MLM and use ensemble methods to increase the performance on original classification problems.
Language
English (en)
Chair
Shantanu Chakrabartty
Committee Members
Ayan Chakrabarti, Arye Nehorai