Abstract

Deep Learning networks are a new type of neural network that discovers important object features. These networks determine features without supervision, and are adept at learning high level abstractions about their data sets. These networks are useful for a variety of tasks, but are difficult to train. This difficulty is compounded when multiple networks are trained in a layered fashion, which results in increased solution complexity as well as increased training time. This paper examines the use of Genetic Algorithms as a training mechanism for Deep Learning networks, with emphasis on training networks with a large number of layers, each of which is trained independently to reduce the computational burden and increase the overall flexibility of the algorithm. This paper covers the implementation of a multilayer deep learning network using a genetic algorithm, including tuning the genetic algorithm, as well as results of experiments involving data compression and object classification. This paper aims to show that a genetic algorithm can be used to train a non trivial deep learning network in place of existing methodologies for network training, and that the features extracted can be used for a variety of real world computational problems.

Library of Congress Subject Headings

Neural networks (Computer science); Genetic algorithms; Machine learning

Publication Date

2012

Document Type

Thesis

Department, Program, or Center

Computer Science (GCCIS)

Advisor

Gaborski, Roger

Comments

Note: imported from RIT’s Digital Media Library running on DSpace to RIT Scholar Works. Physical copy available through RIT's The Wallace Library at: QA76.87 .L36 2012

Campus

RIT – Main Campus

Download

Share

COinS