Bispectrum estimation using a recurrent neural network
Takehiko Ogawa, Yukio Kosugi · Electronics and Communications in Japan (Part III Fundamental Electronic Science) · 2000
Techniques using FFT have been used widely in the past for bispectrum estimation. However, when FFT was used for bispectrum estimation of data with many points, there was a problem of escalated level of calculation, making applications to real problems requiring real-time processing such as image processing difficult. Although the parametric estimation method using AR models by Raghuveer and Nikias and colleagues has been effective in its capability to eliminate Gaussian noise, it has not resolved the problem of the level of calculation associated with parameter estimation. In this study, the case of obtaining the bispectra of continuously changing real waveforms has been considered and a method of estimating bispectra using an AR model network and estimating parameters by the learning of the network is proposed. Although in general a great deal of time is required for the network to learn from random states, the learning of the network is made possible with a small level of calculation by using a method of additional learning. © 2000 Scripta Technica, Electron Comm Jpn Pt 3, 83(10): 91–99, 2000