Robust Rule Based Neural Network Using Arithmetic Fuzzy Inference System

Deep Neural Networks (DNNs) are currently one of the most important research areas of Artificial Intelligence (AI). Various type of DNNs have been proposed to solve practical problems in various fields. However the performance of all these types of DNNs degrades in the presence of feature noise. Expert systems are also a key area of AI that are based on rules. In this work we wish to combine the advantages of these two areas. Here, we present Rule-Based Neural Networks (RBNNs) where each neuron is a Fuzzy Inference System (FIS). RBNN can be trained to learn various regression and classification tasks. It has relatively a few trainable parameters. It is robust to (input) feature noise and it provides a good prediction accuracy even in the presence of large feature noise. The learning capacity of the RBNN can be enhanced by increasing the number of neurons, number of rules and number of hidden layers. The effectiveness of RBNNs is demonstrated by learning real world regression and classification tasks.