著者

宇野 達也 小圷 成一 平田 廣則

出版者

一般社団法人 電気学会

雑誌

電気学会論文誌C(電子・情報・システム部門誌) (ISSN:03854221)

巻号頁・発行日

vol.116, no.10, pp.1183-1187, 1996-09-20 (Released:2008-12-19)

参考文献数

11

---

It is difficult to decide the optimal neural network structure for a particular problem. We propose a solution to this problem, a new constructive learning algorithm based on division of a given learning problem. The proposed method first decomposes the original learning problem into small pieces and constructs a set of small networks which independently learn one of decomposed problems. It constructs a large network which learns the given learning problem by combining the small networks in a bottom-up manner. We demonstrate the efficiency of our learning algorithm by applying it to XOR, 3-bits parity, a non-liner function approximation, and two-spirals problem. Experimental results show that our learning algorithm can construct networks which have higher learning convergence rate and better generalization capability within less computation time than the standard back-propagation algorithms.

著者

宇野 達也 小圷 成一 平田 廣則

出版者

一般社団法人 電気学会

雑誌

電気学会論文誌C(電子・情報・システム部門誌) (ISSN:03854221)

巻号頁・発行日

vol.118, no.3, pp.326-332, 1998-03-01 (Released:2008-12-19)

参考文献数

15

---

We propose a new ANN learning algorithm based on hierarchical clustering of training data. The proposed algorithm first constructs a tree of sub-learning problems by hiearchically clustering given learning patterns in a bottom-up manner and decides a corresponding network structure. The proposed algorithm trains the whole network giving teacher signals of the original learning problem to the output units, and trains sub-networks giving teacher signals of the divided sub-learning problems to the hidden units simultaneously. The hidden units which learn sub-learning problems become feature detectors, which promote the learning of the original learning problem. We demonstrate the advantages of our learning algorithm by solving N-bits parity problems, a non-liner function approximation, iris classification problem, and two-spirals problem. Experimen-tal results show that our learning algorithm obtains better solutions than the standard back-propagation algorithms and one of constructive algorithms in terms of the learning speed and the convergence rate.