Controlling Relations between the Individuality and Collectivity of Neurons and its Application to Self-Organizing Maps

In this paper, we consider a society of neurons where different types of neurons interact with each other. For the first approximation to this society, we suppose two types of neurons, namely, individually and collectively treated neurons. Just as individuality must be in harmony with collectivity in actual societies, individually treated neurons must cooperate with collectively treated neurons as much as possible. We here realize this cooperation by making individually treated neurons as similar to collectively treated neurons as possible. The difference between individually and collectively treated neurons is represented by the Kullback–Leibler divergence. This divergence is minimized using free energy minimization. We applied the method to three problems from the well-known machine learning database, namely wine and protein classification, and the image segmentation problem. In all three problems, we succeeded in producing clearer class structures than those obtainable using the conventional SOM. However, we observed that the fidelity to input patterns deteriorated. For this problem, we found that careful treatment of learning processes were needed to keep fidelity to input patterns at an acceptable level.