| | |
Summary: Supervised Training of a Neural Network for
Classification via Successive Modification of the Training
Data - An Experimental Study*
Mayer Aladjem
Department of Electrical and Computer Engineering
Ben-Gurion University of the Negev, P.O.B. 653, 84105 Beer-Sheva,
Israel, e-mail: aladjem@bguee.bgu.ac.il
Abstract. A method for training of an ML network for classification has been proposed by us
in [3,4]. It searches for the non-linear discriminant functions corresponding to several small
local minima of the objective function. This paper presents a comparative study of our method
and conventional training with random initialization of the weights. Experiments with a
synthetic data set and the data set of an OCR problem are discussed. The results obtained
confirm the efficacy of our method which finds solutions with lower misclassification errors
than does conventional training.
Keywords: Neural networks for classification, auto-associative network, projection pursuit,
structure removal, discriminant analysis, statistical pattern recognition.
1 Introduction
The training of a multi-layer (ML) neural network can be formulated in terms
of the minimization of an error function EML
(w) , which depends on the vector w
|
