Electronic Theses and Dissertations

Identifier

3741

Date

2016

Document Type

Dissertation (Access Restricted)

Degree Name

Doctor of Philosophy

Major

Electrical and Computer Engr

Concentration

Computer Engineering

Committee Chair

Bonny Banerjee

Committee Member

Russell Jerry Deaton

Committee Member

Eddie Jacobs

Committee Member

Madhusudhanan Balasubramanian

Abstract

This dissertation is an investigation of unsupervised algorithms for the problem of feature learning from spatiotemporal data. The algorithms include spherical clustering, sparse coding and non-negative matrix factorization. The algorithms are implemented in a two-layered neural model. When exposed to natural videos, the first and second layer features in the model develop simple and complex cell-like receptive field properties. After learning, the first layer features represent small unoriented filters, localized and oriented Gabor-like edge detector and high frequency gratings. The second layer features become selective to a range of orientations and spatial frequencies, but robust to a wide range of positions. The features learned using sparse coding algorithm are successfully deployed for detecting outliers/saliencies/abnormalities in images, videos and other datasets with various outlier scoring functions. Extensive experimentations with a number of state-of-the-art algorithms on thousands of benchmark datasets revealed that the sparse coding based outlier detection method outperforms traditional method for high-dimensional and difficult-to-comprehend datasets. The features are also used for attention-based object recognition where the environment is considered as partially observable and the series of glimpses are considered as actions. Experiments with several object recognition datasets revealed an interesting trend: accuracy doesn't improve after a certain number of glimpses and sometimes decreases with more glimpses if multiple object categories have similar structure. Finally, using the same algorithms, novel and interesting features are learned from errors with wiring length regularization which can be fitted using a Gaussian derivative model. The derivative order generally increases with decreasing neighborhood.

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