Pattern classification using quadratic neuron: An experimental study

Abstract

In this paper, we present a study done on quadratic neurons to solve the pattern classification problems. The paper compares the classification results obtained by quadratic neural network(QUAD) with normal single and multilayer perceptron(MLP). Examples with randomly generated toy datasets are used for understanding and visualization. The standard datasets such as Iris, MNIST and others are used for extensive comparison and interesting experiments. Different architectures of QUAD neural net has also been tested. Obtained results are better in comparison to conventional multilayer perceptron. This experimental study motivates the authors to study usage of QUAD neurons in deep neural networks.