Use of prior knowledge in classification of similar and structured objects

Wang, Li-Lun

Permalink

https://hdl.handle.net/2142/34196 Copy

Description

Title

Use of prior knowledge in classification of similar and structured objects

Author(s)

Wang, Li-Lun

Issue Date

2012-09-18T21:05:15Z

Director of Research (if dissertation) or Advisor (if thesis)

DeJong, Gerald F.

Doctoral Committee Chair(s)

DeJong, Gerald F.

Committee Member(s)

• Roth, Dan
• Hoiem, Derek W.
• Smaragdis, Paris
• Sun, Qiang

Department of Study

Computer Science

Discipline

Computer Science

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• Machine Learning
• Artificial Intelligence
• Prior Knowledge
• Structured Objects
• Explanation-Based Learning
• Feature Construction
• Model Construction
• Character Recognition
• Classification

Abstract

Statistical machine learning has achieved great success in many fields in the last few decades. However, there remain classification problems that computers still struggle to match human performance. Many such problems share the same properties---large within class variability and complex structure in the examples, which is often true for real world objects. This does not mean lack of information for classification in the examples. On the contrary, there is still a clear pattern in the examples, but hidden behind a many-way covariance structure such that useful information is too dilute for conventional statistical machine learners to pick up. However, if we can exploit the structural nature of the objects and concentrate information about the classification, the problem can become much easier. In this dissertation we propose a framework using prior knowledge about modeling the structures in the examples to concentrate information for classification. The framework is instantiated to the task of classifying pairs of similar offline handwritten Chinese characters. We empirically demonstrate that our proposed framework indeed concentrates useful information for classification and makes the classification problem easier for statistical learning. Our approach advances the state of the art both in offline handwritten character recognition and in machine learning.

Graduation Semester

2012-08

Permalink

http://hdl.handle.net/2142/34196

Copyright and License Information

Copyright 2012 Li-Lun Wang

Owning Collections