Kernel feature detector: extracting kernel features by minimizing α-information

In the present paper, we propose kernel feature detectors for extracting salient features of input patterns. Kernel feature detectors are generated by the information controllers. The information controllers are mainly used to minimize the α-information, difference between Shannon entropy and Renyi entropy, and to generate explicit kernel feature detectors. By minimizing the α-information, a few important features called kernel features are separated from many other unimportant units. We applied our method to two problems: F-H problem and twenty-six alphabet character recognition. In all these cases, we succeeded in extracting kernel features of input patterns, corresponding to our intuition about the input patterns