Feature detection by structural information: complex information control to detect features in neural networks

We propose new information called structural information. The information is introduced to detect the main features, depending on given problems. The structural information is composed of different levels of information. By selecting and controlling the appropriate level of information, we can control internal representations for feature detection. The structural information was applied to an alphabet character recognition problem and a language acquisition problem. In the alphabet character problem we show that different internal representations are generated by changing the structural information. In the language acquisition problem it is shown that generated features and internal representations are easily interpreted.