Improving Artificial Neural Network Classification through Neuron Specialization

In this paper, a novel artificial neural network extension is proposed that draws inspiration from the fact that actual biological brains consist of various types of neurons. The method known as “neuron specialization" consists of training not only the output neurons, but also the hidden neurons to activate for specific classes. This creates neurons that the output neurons can easily rely on and improves the interpretation of the hidden neurons. The first part of this research consists of exploring the neural network model intricacies and performances. Afterwards, the neural network extension is tested using two case studies: a marketing case and an image recognition case. The neural specialization extension is able to achieve significant performance boosts and interpretation improvements for several different neural network architectures.