An experimental evaluation of weightless neural networks for multi-class classification

WiSARD belongs to the class of weightless neural networks, and it is based on a neural model which uses lookup tables to store the function computed by each neuron rather than storing it in weights of neuron connections. WiSARD is characterised by a simple implementation and a fast learning phase due to one-way RAM access/lookup mechanism. WiSARD was originally conceived as a pattern recognition device mainly focusing on image processing. In this work we present a multi-class classification method in machine learning domain based on WiSARD, called WiSARD Classifier. The method uses the same binary encoding scheme to transform multivariable data in the domain of real numbers into binary patterns which are the input to WiSARD. The main contribution of this work is an extensive experimental evaluation of WiSARD's classification capability in comparison to methods from the state-of-the-art. For the purpose we conducted many experiments applying nine well known machine learning methods (including the WiSARD Classifier) to seventy classification problems. Cross-validation accuracies were collected and compared by means of a statistical analysis based on nonparametric tests (Friedman, Friedman Aligned Rank, and Quade test) to prove how the WiSARD Classifier is very close in performance to the best methods available in most popular machine learning libraries.