Fundamentals of Deep Learning and Deep Visualization

Abstract

Recently, replacement of deep learning with neural network is accelerating in the domains to which neural networks have been applied. And there have been many reports that the deep learning contributed to high accuracy and efficiency. However, it has been pointed out that there are some disadvantages for deep learning. In particular, "the reasons for high accuracy and efficiency are not clear" becomes a critical issue. On the other hand, research to explore the mechanism of deep learning is done theoretically and by detailed analysis of internal state of deep learner. The latter representative method is deep visualization. In this paper, we first explain some fundamentals about deep learning based on neural network. And then, by deep visualization, we show that what has been revealed about the reasons for high accuracy and efficiency, and introduce some examples of use of the intermediate output measures.