Simplifying OCR neural networks with oracle learning

Author

Menke, Joshua ; Martinez, Tony

Author_Institution

Dept. of Comput. Sci., Brigham Young Univ., Provo, UT, USA

fYear

2003

fDate

5/17/2003 12:00:00 AM

Firstpage

6

Lastpage

13

Abstract

Often the best model to solve a real world problem is relatively complex. The article presents oracle learning, a method using a larger model as an oracle to train a smaller model on unlabeled data in order to obtain: (1) a simpler acceptable model and (2) improved results over standard training methods on a similarly sized smaller model. In particular, this paper looks at oracle learning as applied to multilayer perceptrons trained using standard backpropagation. For optical character recognition, oracle learning results in an 11.40% average decrease in error over direct training while maintaining 98.95% of the initial oracle accuracy.

Keywords

backpropagation; feedforward neural nets; multilayer perceptrons; optical character recognition; OCR neural network; acceptable model; backpropagation; error decrease; model training; multilayer perceptron; optical character recognition; oracle accuracy; oracle learning; training method; unlabeled data; Artificial neural networks; Backpropagation; Character recognition; Computer science; Design methodology; Multilayer perceptrons; Neural networks; Optical character recognition software; Optical computing; Training data;

fLanguage

English

Publisher

ieee

Conference_Titel

Soft Computing Techniques in Instrumentation, Measurement and Related Applications, 2003. SCIMA 2003. IEEE International Workshop on

Print_ISBN

0-7803-7711-7

Type

conf

DOI

10.1109/SCIMA.2003.1215923

Filename

1215923