Handwritten Numeral Recognition Using Neural Networks

Firas Hamze
EE368, Stanford University

27 May, 2000

Abstract:

We explore a neural network approach to handwritten digit recognition. A theoretical explanation of feed-forward networks is presented. Several fully-connected, feed-forward neural nets of varying size are trained on images obtained from the NIST online database. Performance of the systems is analyzed and compared to that of Principal Component Analysis, which is also given a brief theoretical outline. The best-performing system was a 50-unit network, which had a recognition error of 6.95% on the test data. PCA achieved a best error rate of 14%

• Introduction
• The Handwritten Data Set
• Feed-Forward Neural Network Method
  • Neural Network Principles
  • Simulation Details and Results
• Principal Component Analysis(PCA)
  • PCA Principles
  • Simulation Details and Results
• Conclusion
• References
• About this document ...