An Integrated Perceptron Network and Learning Algorithm for Multi-Class Patterns Recognition

A new integrated perceptron network and the related learning algorithm is proposed. The network is made up of three layers, the weights of the middle layer to the output layer being +1's or -1's correspondingly. The weights of the input layer to the middle layer are obtained by learning, and the learning process of each neuron in the middle layer is completed individually. Furthermore, the algorithm is stopped within finite steps. Upon termination of the algorithm, the correct weight must be found for a linearly separable multiclass pattern recognition. If there exist some patterns that cannot be recognized, then the set of patterns is a nonlinearly separable multiclass pattern recognition problem. Simulation results of the recognition of digits show the effectiveness of the proposed network model and algorithm.