Modification of hard-limiting multilayer neural networks for confidence evaluation

The central theme of this paper is to overcome the inability of feedforward neural networks with hard limiting units to provide confidence evaluation. We consider a Madaline architecture for a 2-group classification problem and concentrate on the probability density function for the neural activation of the first-layer units. As the following layers perform a Boolean table, the expectation value of the output is determined, utilizing the probability of a pattern to perform a definite binary input for the Boolean table. The Madaline architecture can be modified to the introduced Σ-Π-Σ network, which evaluates the expectation value. Several assumptions on the distribution of the neural activation lead to a clear and simple architecture, which is applied to an OCR problem.