Dependent input neuron selection in contradiction resolution

Abstract

In this paper, we propose a new type of information theoretic method called “dependent input neuron selection” in the framework of contradiction resolution. In contradiction resolution, a neuron fires without considering other neurons (self-evaluation), and at the same time the neuron's firing rate is determined by other neurons (outer-evaluation). If there exists contradiction between self and outer-evaluation, the contradiction should be reduced as much as possible. Roughly speaking, outer-evaluation corresponds to cooperation between neurons in the self-organizing maps. Thus, contradiction resolution can be applied to the production of self-organizing maps. In this contradiction resolution, we introduce dependent input neuron selection. The importance of neurons is determined by the degree of matching between neurons. A limited number of best-matching input neurons participate in processing input patterns. We applied the method to the CO2 production. Experimental results showed that prediction performance was much improved by choosing the appropriate number of input neurons. In addition, better prediction performance was accompanied by reasonably small quantization and topographic errors. The results suggest a possibility of contradiction resolution to produce networks with higher prediction performance and better topological properties.