Evaluating Binary Encoding Techniques for WiSARD

Abstract

Many weightless neural networks, such as WiSARD, are RAM-based classifiers that receive binary data as input. In order to convert raw data into binary input, several techniques are applicable. This work evaluates the impact of some of these binarization techniques on the accuracy of two types of classifiers: WiSARD model and WiSARD with bleaching mechanism. The binary encoding techniques explored were: (i) thermometer, (ii) threshold, (iii) local threshold, (iv) Marr-Hildreth filter, and (v) Laplacian filter. The MNIST digit dataset was used to compare the accuracy obtained by each encoding technique. Results showed a difference of more than 20% in the accuracy due to the choice of encoding approach.