查詢結果分析
來源資料
頁籤選單縮合
| 題 名 | Learning Multilayer Potts Perceptrons for Blind Deconvolution=學習多層波茲認知機解隱蔽式反迴旋問題 |
|---|---|
| 作 者 | 吳建銘; 吳俊樟; 連崇順; | 書刊名 | 中國統計學報 |
| 卷 期 | 52:1 2014.03[民103.03] |
| 頁 次 | 頁10-23 |
| 分類號 | 319.9 |
| 關鍵詞 | 迴旋模式; 多層認知機; 多層波茲認知機; 隱蔽式等量化; 隱蔽式反迴旋; 隱藏式馬可夫模式; Convolution model; Multilayer perceptrons; Multilayer Potts perceptrons; Blind equalization; Blind deconvolution; Hidden Markov model; |
| 語 文 | 英文(English) |
| 中文摘要 | 本文提出以萃取非線性遞迴結構為目標的 MLPotts (Multilayer Potts Peceptrons,多層波茲認知機 )學習分析方法 ,解決隱蔽式線性迴旋建模及反迴旋問題。假設給定的單一頻道觀察值是由獨立訊號源透過線性迴旋結構所形成的時間序列,在此前提下,本研究使用 MLPotts學習法萃取內嵌於單一頻道訊號的非線性遞迴結構,並且估計未知的線性迴旋結構以回復原始獨立訊號源。藉由以萃取非線性遞迴結構為目標的 MLPotts學習分析,可將所給定的單一頻道訊號分解為兩個部份,分別代表獨立訊號源經由最佳非線性迴旋所形成的可預測成份及剩餘的非決定性成份 (non-deterministic component)。後者的訊號觀察值具備時間獨立特性 ,可以視為獨立訊號源的有效估計。根據所估計的獨立訊號源,可進一步重構未知的迴旋結構。數值模擬顯示,以萃取非線性遞迴結構為目標的 MLPotts學習分析可以有效於解決隱蔽式線性迴旋建模及反迴旋問題,並且可應用於解決隱蔽式非線性迴旋建模問題。 |
| 英文摘要 | This paper explores learning nonlinear recursions for blind linear convolution modeling and deconvolution. Under the assumption that given single-channel observations are linear convolution of independent temporal instances, this work employes MLPotts(Multilayer Potts perceptrons) learning to extract nonlinear recursion embedded within single-channel observations, estimate unknown linear convolution and recover the original independent source. By MLPotts learning of nonlinear recursions given single-channel observations are decomposed to two components, respectively the predictable component of proceeding temporal instances through optimal nonlinear recursion and the residual non-deterministic component. The instance in the latter is independent of predecessors and can be regarded as an estimation to the independent source. Based on the attained independent source, the unknown convolution structure could be further reconstructed. By numerical simulations, it is shown that learning nonlinear recursion is effective for blind linear convolution modeling and deconvolution, and is potential for blind nonlinear convolution modeling. |
本系統中英文摘要資訊取自各篇刊載內容。