Two phases based training method for designing codewords for a set of perceptrons with each perceptron having multi-pulse type activation function

This paper proposes a two phases-based training method to design the codewords to map the cluster indices of the input feature vectors to the outputs of the new perceptrons with the multi-pulse type activation functions. Our proposed method is applied to classify two types of the tachycardias. First, the total number of the new perceptrons is initialized as the dimensions of the input feature vectors. Next, a set of new perceptrons with each new perceptron having a single pulse type activation function is designed. Then, the new perceptrons with the multi-pulse type activation function are designed based on those new perceptrons with the single pulse type activation function. After that, the codewords are assigned according to the outputs of the new perceptrons with the multi-pulse type activation functions. Finally, a condition on the codewords is checked. The significance of this work is to guarantee to achieve the no classification error efficiently through using more than one new perceptron with the multi-pulse type activation if the feature space can be linearly partitioned into the multiple clusters. The computer numerical simulation results show that our proposed method outperforms the conventional perceptrons with the sign type activation function.