Improved MLP Learning via Orthogonal Bipolar Target Vectors

Abstract

In this paper, we present an approach to improve "Multilayer Perceptron" (MLP) learning by adopting orthogonal bipolar vectors as expectation values. These vectors differ from conventional vectors in two main ways. First, since they are larger than conventional ones, they use more synaptic connections among hidden and output neurons. Second, since they are orthogonal vectors with bipolar representation, the Euclidean distance for these vectors (in a Euclidean space Rⁿ) increases when their number (n) of components increases. We present an algorithm for constructing orthogonal bipolar vectors used as target vectors. These target vectors are used in experiments for training by backpropagation algorithm with an MLP model that classifies characters extracted from degraded license plate images. Experimental results are obtained by using both orthogonal bipolar and conventional target vectors. Comparisons of results lead to conclusions that recognition of degraded characters are considerably improved with the proposed expectation values for MLP learning.