GUACAMOLE: A new paradigm for unsupervised competitive learning

Abstract

In this paper we present a new model of unsupervised learning called the General Unsupervised Adaptive Classification Algorithm for Modular Organization of Learning Evolution (GUACAMOLE). This system is based on the competition among a certain number of Auto-Associative ANNs (local experts), which, in the context of pattern recognition tasks, develop a specialized knowledge of a certain, specific class, and are subsequently selected by the system on a competitive basis by looking not at error minimization, but rather, at the closeness of errors in the Training and Testing phases - i.e., the chosen expert is the one that is least 'surprised' by reviewing that specific record. We prove on two benchmarks of very different nature that GUACAMOLE delivers superior performance with respect to a wide and diversified battery of competitor systems, both supervised and unsupervised, and conclude that this particular type of unsupervised learning can provide the basis for a new paradigm where unsupervised learning is taken as the natural reference and in which contexts where supervised learning was commonly believed to be the reference option.