Feature extraction by uniform structure threshold logic networks

We propose a uniform structure threshold logic network(USTN) with multi-inputs and multi-outputs, which can extract important features from time-series multi-demensional signals. In this paper, first, we examine the fundamental characteristics of USTN, and verify what feature can be extracted by the USTN. Next, we apply the USTN to extract the important features from soectral pattern of speech or hand-written Chinese character pattern. As a result, the USTN can extract formant, which is an important feature of speech, and can extract vertical, horizontal, 45° and -45° components of character pattern. Finally, using the USTN, there is a probability of shortening recognition time and making the recognition system as an LSI chip.