Electronic Theses and Dissertations

Identifier

995

Date

2013

Date of Award

11-27-2013

Document Type

Thesis

Degree Name

Master of Science

Major

Electrical and Computer Engr

Concentration

Computer Engineering

Committee Chair

Bonny Banerjee

Committee Member

Robert Kozma

Committee Member

Charles Blaha

Committee Member

Eddie Jacobs

Abstract

The success of many tasks depends on good feature representation which is often domain-specific and hand-crafted requiring substantial human effort. Such feature representation is not general, i.e. unsuitable for even the same task across multiple domains, let alone different tasks.To address these issues, a multilayered convergent neural architecture is presented for learning from repeating spatially and temporally coincident patterns in data at multiple levels of abstraction. The bottom-up weights in each layer are learned to encode a hierarchy of overcomplete and sparse feature dictionaries from space- and time-varying sensory data. Two algorithms are investigated: recursive layer-by-layer spherical clustering and sparse coding to learn feature hierarchies. The model scales to full-sized high-dimensional input data and to an arbitrary number of layers thereby having the capability to capture features at any level of abstraction. The model learns features that correspond to objects in higher layers and object-parts in lower layers.Learning features invariant to arbitrary transformations in the data is a requirement for any effective and efficient representation system, biological or artificial. Each layer in the proposed network is composed of simple and complex sublayers motivated by the layered organization of the primary visual cortex. When exposed to natural videos, the model develops simple and complex cell-like receptive field properties. The model can predict by learning lateral connections among the simple sublayer neurons. A topographic map to their spatial features emerges by minimizing the wiring length simultaneously with feature learning.The model is general-purpose, unsupervised and online. Operations in each layer of the model can be implemented in parallelized hardware, making it very efficient for real world applications.

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.

Share

COinS