The present paper shows how to solve the problem of vanishing information in potential mutual information maximization. We have previously developed a new information-theoretic method called "potential learning" which aims to extract the most important features through simplified information maximization. However, one of the major problems is that the potential effect diminishes considerably in the course of learning and it becomes impossible to take into account the potentiality in learning. To solve this problem, we here introduce repeated information maximization. To enhance the processes of information maximization, the method forces the potentiality to be assimilated in learning every time it becomes ineffective. The method was applied to the on-line article popularity data set to estimate the popularity of articles. To demonstrate the effectiveness of the method, the number of hidden neurons was made excessively large and set to 50. The results show that the potentiality information maximization could increase mutual information even with 50 hidden neurons, and lead to improved generalization performance. In addition, simplified representations could be obtained for better interpretation and generalization.
展开▼
