※一部利用できない機能があります
| 1.
論文 |
論文
|
Jansen, Boris ; Nakayama, Kenji
概要:
金沢大学大学院自然科学研究科情報システム<br />Over the years, many improvements and refinements to the backpropagation learning algorithm ha
…
ve been reported. In this paper, a new adaptive penalty-based learning extension for the backpropagation learning algorithm and its variants is proposed. The new method initially puts pressure on artificial neural networks in order to get all outputs for all training patterns into the correct half of the output range, instead of mainly focusing on minimizing the difference between the target and actual output values. The upper bound of the penalty values is also controlled. The technique is easy to implement and computationally inexpensive. In this study, the new approach is applied to the backpropagation learning algorithm as well as the RPROP learning algorithm. The superiority of the new proposed method is demonstrated though many simulations. By applying the extension, the percentage of successful runs can be greatly increased and the average number of epochs to convergence can be well reduced on various problem instances. The behavior of the penalty values during training is also analyzed and their active role within the learning process is confirmed. Copyright © 2006 The Institute of Electronics, Information and Communication Engineers.
続きを見る
|
||||
| 2.
論文 |
論文
|
Katou, Haruo ; Nakayama, Kenji ; Hirano, Akihiro
概要:
金沢大学理工研究域 電子情報学系<br />本稿では,信号源の数がセンサ数より多いオーバーコンプリート・ブラインド信号源分離においてフィードバック形構成法と分離回路の学習アルゴリズムを提案する.まず,1巡目で信号源の分離を行う.ここでは,一
…
つの信号源が複数の出力に含まれないことを分離の条件とする.このための学習法を提案している.センサ数を信号源数の約半分以上とすることにより,1巡目の信号源分離で少なくとも1個の出力に単一信号源を分離できる.この出力を単一信号源の特徴を利用して検出する.更に,この出力をフィードバックして観測信号からキャンセルすることにより,等価的に信号源の数を低減する.当該出力と混合過程の情報を使ってキャンセルする他,観測信号と当該出力のヒストグラムを使ってキャンセルする方法を組み合わせることにより,条件不足の問題を解消する.2巡目では,観測信号に含まれる信号源が1個少ない状態で1巡目と同じ処理を行う.このように,提案法では,観測信号における信号源の数を1個ずつ減らしながら信号源分離を繰り返す.信号源として音声を用いたシミュレーションにより,従来法との比較を行い,提案方法の有効性を確認している. A feedback structure and its learning algorithm are proposed for overcomplete blind source separation, where the number of sources is larger than that of sensors. In the first phase, the signal sources are separated so as to satisfy the condition, under which one signal source is not included in different outputs. By setting the number of the sensors to be more than a half of the sources, at least one output includes a single source. This output is selected based on property of a single source. Furthermore, this output is fed back and cancelled from observed signals. The selected single source, information about a mixing process and histogram of the observations and the separated source are used for this cancellation. In the second phase, the same process is carried out by using the modified observations, in which the source, separated in the first phase, is cancelled. Like this, in the proposed method, the source separation is repeated by reducing the number of equivalent sources. Simulations using speech signals demonstrate usefulness of the proposed method compared to the conventional methods.
続きを見る
|
||||
| 3.
論文 |
論文
|
Jansen, Boris ; Nakayama, Kenji
概要:
金沢大学理工研究域 電子情報学系<br />Over the years, many improvements and refinements to the backpropagation learning algorithm have b
…
een reported. In this paper, a new adaptive penalty-based learning extension for the backpropagation learning algorithm and its variants is proposed. The new method initially puts pressure on artificial neural networks in order to get all outputs for all training patterns into the correct half of the output range, instead of mainly focusing on minimizing the difference between the target and actual output values. The upper bound of the penalty values is also controlled. The technique is easy to implement and computationally inexpensive. In this study, the new approach is applied to the backpropagation learning algorithm as well as the RPROP learning algorithm. The superiority of the new proposed method is demonstrated though many simulations. By applying the extension, the percentage of successful runs can be greatly increased and the average number of epochs to convergence can be well reduced on various problem instances. The behavior of the penalty values during training is also analyzed and their active role within the learning process is confirmed. Copyright © 2006 The Institute of Electronics, Information and Communication Engineers.
続きを見る
|
||||
| 4.
論文 |
論文
|
Horita, Akihide ; Nakayama, Kenji ; Hirano, Akihiro ; Dejima, Yasuhiro
概要:
金沢大学理工研究域 電子情報学系<br />畳み込み混合過程におけるフィードフォワード(FF-)形ブラインドソースセパレーション(BSS)では自由度が存在するため信号歪みが生じる.我々は,信号源-センサーが2チャンネルの場合において,完全分
…
離と信号無歪みの条件を制約条件として課す信号歪み抑制学習アルゴリズムを時間領域,周波数領域のFF-BSSに対して提案してきた.本稿では,信号歪み抑制の制約条件を多チャンネルに拡張し,かつ,計算の複雑さを軽減するために制約条件を近似する方式を提案する.音声を用いたコンピュータシミュレーションによってその近似制約方式と厳密制約方式がほぼ同等の分離性能と信号歪み抑制が得られることを確認した.また,3チャンネルにおいても,従来方式より特性が改善されることを確認した. Feed-forward Blind Source Separation (FF-BSS) systems have some degree of freedom in the solution space, and signal distortion is likely to occur in convolutive mixtures. Previously, a condition for complete separation and distortion free has been derived for 2-channel FF-BSS. This condition has been applied to the learning algorithms as a distortion free constraint in both the time and frequency domains. In this paper, the condition is further extended to multiple channel FF-BSSs. This condition requires the a high computational complexity to be applied to the learning process as a constraint. An approximate constraint is proposed in order to relax the high computational load. In comparison with the original constraint, computer simulations have demonstrated that the approximation can obtain similar performances with respect to source separation as well as signal distortion using speech signals. Furthermore, the performances can be improved compared to the conventionals for three channels.
続きを見る
|
||||
| 5.
論文 |
論文
|
中山, 謙二 ; Nakayama, Kenji
概要:
金沢大学自然科学研究科<br />信号源やその伝達・混合過程に関する事前情報がなく,それらの統計的な性質のみを仮定して信号源を分離するブラインド形信号源分離(BSS)に関して研究を行った.特に,伝達・混合過程に非線形を含む場合は信号源の混合
…
も複雑になり,それらを分離することは一般に難しい.本研究では,信号源の高次項を含む信号群を分離する過程と,その後に高次項を抑制する線形化の過程を縦続接続する方法を提案した.また,この構成法に対する学習法を新たに提案した.音声の信号源が2個の場合と3個の場合についてシミュレーションを行った.観測信号に比べて,干渉成分と非線形成分(高次項)は約20dB程度減衰しており,有効性が確認できた.本方法では,信号源の数より多い観測センサーが必要とされる.これに関して,非線形成分の大きさと必要とされる観測センサー数の関係について解析し,実際の応用場面では,センサー数を低減できることを示した.さらに,線形化→信号群分離→線形化というサンドウィッチ構造を提案し,少ないセンサー数で良好な分離特性を得た.ブラインド形信号源分離では,分離回路において信号が歪むという問題がある.これに関して,フィードフォワード形BSS(FF-BSS)とフィードバック形BSS(FB-BSS)について解析を行い,信号歪みのメカニズムを始めて明らかにした.FF-BSSは分離回路の自由度が高く,信号歪みが生じる.一方,FB-BSSは観測信号から信号歪みを生じることなく,分離が可能であることが明らかになった.さらに,本研究では,FF-BSSにおいても信号歪みを抑制するための制約条件を付加した学習法を提案し,時間領域における学習で従来法に比べて信号歪みを抑えることができた.以上,本研究では,非線形混合過程におけるBSSに対して新しい方法を提案し,さらに,従来行われていなかった信号歪みの問題を解析し,新しい学習法を提案した.これらにより,BSSをより現実的な場面に応用することが可能になり,また,分離後においても良好な音質が保証される.<br />In practical applications of BSS, processes of generating mixing and sensing signals include nonlinearity, caused by loud speakers, microphones, amplifiers and so on. BSS, cascading a signal group separation block and a linearization block has been proposed for low-order nonlinear mixtures. In the separation block, the signal sources are separated into each group, including its high-order components. The high-order components are further suppressed through the linearization block.In this report, separation performance of the nonlinear BSS is analyzed from several view points. The number of the sensors is increased from that of the signal sources in order to cancel the interference. Moreover, the interference components is decided by a ratio of the nonlinear and the linear components. A relation between the ratio of the components and the number of the sensors is analyzed. The number of the sensors can be reduced when the ratio of the nonlinearity is small. And a Cascade Form BSS Connecting Linearization and Source Separation and Linearization is analyzed.Next, effects of the initial guess of the separation matrix is analyzed. The training was carried out using 50 independent random initial guess, and good separation is obtained by a 25% probability. Moreover, effect of including 3rd-order terms is analyzed. When the 3rd-order term is under 10%, good separation performance can be obtained.<br />研究課題/領域番号:15560323, 研究期間(年度):2003 – 2004<br />出典：「非線形たたみ込み混合過程におけるブラインド形信号源分離・推定法」研究成果報告書 課題番号15560323（KAKEN：科学研究費助成事業データベース（国立情報学研究所））（https://kaken.nii.ac.jp/ja/report/KAKENHI-PROJECT-15560323/155603232004kenkyu_seika_hokoku_gaiyo/)を加工して作成
続きを見る
|
