講演抄録/キーワード |
講演名 |
2005-06-27 10:45
Independent Component Analysis of Signals using Local Exponential Nonlinearities ○Muhammad Tufail・Masahide Abe・Masayuki Kawamata(Tohoku Univ.) |
抄録 |
(和) |
In this paper we propose exponential type nonlinearities in order to blindly separate instantaneous mixtures of signals with any symmetric distributions (using the relative gradient algorithm). These nonlinearities are applied locally, around zero, to avoid degradation of the separating algorithm in the presence of outliers. Such an approach is particularly effective in cases where the sources consist of super-Gaussian signals that inherently contain large data values. The optimal size of the neighborhood, for both continuous and discrete distributions, is obtained by examining the local stability conditions of the relative gradient algorithm. Finally, some computer simulations are presented to demonstrate the superior performance of the proposed idea. |
(英) |
In this paper we propose exponential type nonlinearities in order to blindly separate instantaneous mixtures of signals with any symmetric distributions (using the relative gradient algorithm). These nonlinearities are applied locally, around zero, to avoid degradation of the separating algorithm in the presence of outliers. Such an approach is particularly effective in cases where the sources consist of super-Gaussian signals that inherently contain large data values. The optimal size of the neighborhood, for both continuous and discrete distributions, is obtained by examining the local stability conditions of the relative gradient algorithm. Finally, some computer simulations are presented to demonstrate the superior performance of the proposed idea. |
キーワード |
(和) |
Blind Source Separation / Independent Component Analysis / Relative Gradient Algorithm / / / / / |
(英) |
Blind Source Separation / Independent Component Analysis / Relative Gradient Algorithm / / / / / |
文献情報 |
信学技報, vol. 105, no. 149, SIP2005-28, pp. 19-23, 2005年6月. |
資料番号 |
SIP2005-28 |
発行日 |
2005-06-20 (CAS, VLD, SIP) |
ISSN |
Print edition: ISSN 0913-5685 |
PDFダウンロード |
|
|