Extending WiSARD to perform ensemble learning, regression, multi-label, and multi-modal tasks

Abstract

Wilkie, Stonham and Aleksander’s Recognition Device (WiSARD) is a machine learning model that does not require any kind of error minimization technique to learn patterns. This model uses RAMs as neurons, requiring only one process of writing in memory in its training phase and reading in memory in its classification phase. WiSARD is a weightless artificial neural network, a kind of model that has been successfully used in cognitive architectures and artificial consciousness research. This thesis proposes several extensions for this model in order to create the necessary components for a future WiSARD-based emotion-drive cognitive architecture, The contributions of this work include two WiSARD-based multi-label classification systems, five new types of ensembles using both WiSARD, as well as its ClusWiSARD extension, two WiSARD-based non-parametric regression models, one WiSARDbased logistic regression model, and a weightless multi-modal empathy prediction system. This thesis uses a map-based WiSARD implementation, instantiating only the memory locations that were actually trained. All models and systems created for this thesis have been compared with state-of-the-art, being competitive in some cases, while preserving all the qualities of the canonical WiSARD.