Electronic Theses and Dissertations

Identifier

4833

Date

2016

Document Type

Thesis (Campus Access Only)

Degree Name

Master of Science

Major

Electrical and Computer Engr

Concentration

Computer Engineering

Committee Chair

Bonny Banerjee

Committee Member

Dale Bowman

Committee Member

Eddie Jacobs

Committee Member

Madhusudhanan Balasubramanian

Abstract

Unsupervised feature learning is one of the key components of machine learningand articial intelligence. Learning features from high dimensional streaming data isan important and dicult problem which is incorporated with number of challenges.Moreover, feature learning algorithms need to be evaluated and generalized for timeseries with dierent patterns and components. A detailed study is needed to clarifywhen simple algorithms fail to learn features and whether we need more complicatedmethods.In this thesis, we show that the systematic way to learn meaningful featuresfrom time-series is by using convolutional or shift-invariant versions of unsupervisedfeature learning. We experimentally compare the shift-invariant versions of clustering,sparse coding and non-negative matrix factorization algorithms for: reconstruction,noise separation, prediction, classication and simulating auditory lters from acousticsignals. The results show that the most ecient and highly scalable clustering algorithmwith a simple modication in inference and learning phase is able to produce meaningfulresults. Clustering features are also comparable with sparse coding and non-negativematrix factorization in most of the tasks (e.g. classication) and even more successful insome (e.g. prediction). Shift invariant sparse coding is also used on a novel application,inferring hearing loss from speech signal and produced promising results.Performance of algorithms with regard to some important factors such as: timeseries components, number of features and size of receptive eld is also analyzed. Theresults show that there is a signicant positive correlation between performance of clusteringwith degree of trend, frequency skewness, frequency kurtosis and serial correlationof data, whereas, the correlation is negative in the case of dataset average bandwidth.Performance of shift invariant sparse coding is aected by frequency skewness, frequencykurtosis and serial correlation of data. Non-Negative matrix factorization is influenced by data characteristics same as clustering.

Comments

Data is provided by the student.

Library Comment

Dissertation or thesis originally submitted to the local University of Memphis Electronic Theses & dissertation (ETD) Repository.

Download

Share

COinS