Abstract
In this paper, some input schemes, output representation schemes and learning factors are examined to improve generalization in feedforward three layer Artificial Neural Network with application to handwritten numeral recognition.
The paper focuses on simple experiments with the raw input i.e. no features are extracted. Input schemes discussed includes training with normal training set and extended training set. Output representations include normal 4 bit and 10 bit coding. Experiments are conducted with different learning rates and techniques of ‘weight smoothing’ is used. Lastly it is observed that the network size can be curtailed by removing some of the input neurons which are not contributing much with very little effect on generalization ability of the network.
Additional information
Notes on contributors
S R Gengaje
Sachin R Gengaje is currently working as a Head of the Electronics Department in Walchand Institute of Technology, Solapur. He completed his BE (Electronics) from Walchand Institute of Technology, Solapur in 1990 and ME (Electronics) from Walachand College of Engineering, Sangli in 1995. Presently he is doing PhD in the area of Artificial Neural Networks.
His area of interest are Artificial Neural Networks, Image Processing, Expert Systems and Microprocessors. Sachin R Gengaje is a member of IETE and ISTE.
A R Yardi
A R Yardi was born on 27 August 1949. He took his degree in BE (Electronics & Telecom) from College of Engineering, Pune in 1971 and then degree in M Tech from Indian Institute of Technology, Mumbai in 1975. Since then he is working with Walchand College of Engineering, Sangli in Department of Electronics. At present he is working as a Professor in Electronics. His fields of interest are Image Processing, Pattern Recognition and Artificial Neural Networks. He has till now guided thirty dissertations at ME (Electronics). He is a Life Member of Indian Society for Technical Education.