著者

佐藤 佳代 吉川 暢宏 楊 元稙 中埜 良昭

出版者

日本計算工学会

雑誌

日本計算工学会論文集 (ISSN:13478826)

巻号頁・発行日

vol.1999, pp.19990025-19990025, 1999 (Released:2004-11-11)

参考文献数

10

---

A new learning algorithm named whole learning algorithm is proposed for the feedforward neural network. Strictly speaking, the learning of the feedforward neural network is a kind of multi-objective optimization problem to minimize the errors of outputs for all the learning data sets with respect to the amount of weight modification. All the learning data sets are simultaneously taken into account to constitute the governing equation of the weight modification, which is formulated as linear simultaneous equations with rectangular matrix of coefficients in the proposed algorithm. The solution of the equation is determined by means of the Moore-Penrose generalized inverse to deal with the rectangular matrix. The efficiency of the proposed algorithm is demonstrated through the problem to learn the nolinear behavior described by the Ramberg-Osgood model. The applicability of the proposed algorithm is investigated in problem to learn the earthquake response of RC members.

著者

永野マドセン 泰子 山元 淑乃 楊 元 Nagano Madsen Yasuko Yamamoto Yoshino Yang Yuan

出版者

琉球大学留学生センター

雑誌

留学生教育 : 琉球大学留学生センター紀要 (ISSN:13488368)

巻号頁・発行日

no.10, pp.17-34, 2013-03

---

中国語雲南方言の単一話者による日本語音声を分析し、その全体像を把握した。日本人が聞いて不自然でかつ学習効果が少ないと思われる問題点は「子音の強さ」、濁音が清音になる、および「ら行」の音であった。これらについては音響分析により日本人母語話者との差異を指摘した。母音については、狭母音と広母音の差が少ない中国語の特徴がそっくり日本語の母音のパターンに移されていることが観察された。これらの結果から、雲南方言でも北京方言同様、子音の問題が大きいことが明らかになった。Japanese speech sounds produced by a single Chinese speaker was analyzed. The results showed three areas that are most problematic: obstruents being too strong, voiced obstruent become voiceless, and the realization of /r/ is different. For these, spectrographic analyses showed the difference between a native Japanese speaker and the Chinese learner. The Chinese vowel formant pattern was nearly directly transferred to the Japanese vowel pattern and little effect of learning was observed.