Comprehensibility maximization and humanly comprehensible representations

Abstract

In this paper, we propose a new information-theoretic method to measure the comprehensibility of network configurations in competitive learning. Comprehensibility is supposed to be measured by information contained in components in competitive networks. Thus, the increase in information corresponds to the increase in comprehensibility of network configurations. One of the most important characteristics of the method is that parameters can be explicitly determined so as to produce a state where the different types of comprehensibility can be mutually increased. We applied the method to two problems, namely an artificial data set and the ionosphere data from the well-known machine learning database. In both problems, we showed that improved performance could be obtained in terms of all types of comprehensibility and quantization errors. For the topographic errors, we found that updating connection weights prevented them from increasing. Then, the optimal values of comprehensibility could be explicitly determined, and clearer class boundaries were generated.

Keywords:

• neural networks
• SOM
• competitive learning
• information-theoretic
• mutual information

Acknowledgements

The author is very grateful to the editor and Mitali Das for her valuable comments and suggestions.

Notes

1. http://archive.ics.uci.edu/ml/