Self-organizing Maps of Artificial Neural Classifiers - A Brain-Like Pin Factory

Abstract

Most machine learning algorithms are based on the formulation of an optimization problem using a global loss criterion. The essence of this formulation is a top-down engineering thinking that might have some limitations on the way towards a general artificial intelligence. In contrast, self-organizing maps use cooperative and competitive bottom-up rules to generate low-dimensional representations of complex input data. Following similar rules to SOMs, we develop a self-organization approach for a system of classifiers that combines top-down and bottom-up principles in a machine learning system. We believe that such a combination will overcome the limitations with respect to autonomous learning, robustness and self-repair that exist for pure top-down systems. Here we present a preliminary study using simple subsystems with limited learning capacities. As proof of principle, we study a network of simple artificial neural classifiers on the MNIST data set. Each classifier is able to recognize only one single digit. We demonstrate that upon training, the different classifiers are able to specialize their learning for a particular digit and cluster according to the digits. The entire system is capable of recognizing all digits and demonstrates the feasibility of combining bottom-up and top-down principles to solve a more complex task, while exhibiting strong spontaneous organization and robustness.