A reduced and comprehensible polynomial neural network for classification

It has been found that in solving classification task, the polynomial neural network (PNN) needs more computation time, as the partial descriptions (the heart of PNN) in each layer grow very fast. At the same time the complexity of the network also increases as the number of layers increases. In this context we propose a reduced and comprehensible polynomial neural network (RCPNN) for the task of classification for which partial descriptions have been developed only for a single layer of the PNN architecture and the output of these partial descriptions along with the features are fed to the output layer of the RCPNN having only one neuron. The weights between hidden layer and output layer have optimized by two different methods such as gradient descent and particle swarm optimization (PSO). A comparative performance in terms of computational cost and accuracy of PSO trained RCPNN and non-PSO (i.e. gradient descent) trained RCPNN with PNN has been given to prove the same. Our experimental results show that the performance in terms of cost and accuracy of the proposed RCPNN trained with PSO and gradient decent is more efficient than the PNN model.