Abstract
A neural network is developed to mimic a human brain. The neural network consists of units and links that connect between units. Various types of neural networks are categorized into two classes: (1) back-propagation hierarchical neural network and (2) mutual-connected neural network. Generally speaking, it is hard to fix the number of units to build a neural network for solving problems. So the number of units is decided on the basis of experts' experience. In this paper, we explain a learning method how to decide the structure of a neural network for problems. The learning method is named structural learning. Even if we give a sufficient number of units, the optimal structure will be decided in the process of learning. The objective of the paper is to explain the structural learning of both hierarchical and mutual connecting neural networks. Both networks obtained and showed the sufficiently good results. In the stock forecast by a general neural network, the operation and the system cost are very large because a lot of numbers of hidden layer units in the network are used. This research tried the optimization of the network by the structured learning, and evaluated the practicality.
Original language | English |
---|---|
Title of host publication | Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) |
Pages | 231-240 |
Number of pages | 10 |
Volume | 4529 LNAI |
Publication status | Published - 2007 |
Event | 12th International Fuzzy Systems Association World Congress, IFSA 2007 - Cancun Duration: 2007 Jun 18 → 2007 Jun 21 |
Publication series
Name | Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) |
---|---|
Volume | 4529 LNAI |
ISSN (Print) | 03029743 |
ISSN (Electronic) | 16113349 |
Other
Other | 12th International Fuzzy Systems Association World Congress, IFSA 2007 |
---|---|
City | Cancun |
Period | 07/6/18 → 07/6/21 |
Fingerprint
ASJC Scopus subject areas
- Computer Science(all)
- Biochemistry, Genetics and Molecular Biology(all)
- Theoretical Computer Science
Cite this
New perspective for structural learning method of neural networks. / Watada, Junzo.
Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics). Vol. 4529 LNAI 2007. p. 231-240 (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics); Vol. 4529 LNAI).Research output: Chapter in Book/Report/Conference proceeding › Conference contribution
}
TY - GEN
T1 - New perspective for structural learning method of neural networks
AU - Watada, Junzo
PY - 2007
Y1 - 2007
N2 - A neural network is developed to mimic a human brain. The neural network consists of units and links that connect between units. Various types of neural networks are categorized into two classes: (1) back-propagation hierarchical neural network and (2) mutual-connected neural network. Generally speaking, it is hard to fix the number of units to build a neural network for solving problems. So the number of units is decided on the basis of experts' experience. In this paper, we explain a learning method how to decide the structure of a neural network for problems. The learning method is named structural learning. Even if we give a sufficient number of units, the optimal structure will be decided in the process of learning. The objective of the paper is to explain the structural learning of both hierarchical and mutual connecting neural networks. Both networks obtained and showed the sufficiently good results. In the stock forecast by a general neural network, the operation and the system cost are very large because a lot of numbers of hidden layer units in the network are used. This research tried the optimization of the network by the structured learning, and evaluated the practicality.
AB - A neural network is developed to mimic a human brain. The neural network consists of units and links that connect between units. Various types of neural networks are categorized into two classes: (1) back-propagation hierarchical neural network and (2) mutual-connected neural network. Generally speaking, it is hard to fix the number of units to build a neural network for solving problems. So the number of units is decided on the basis of experts' experience. In this paper, we explain a learning method how to decide the structure of a neural network for problems. The learning method is named structural learning. Even if we give a sufficient number of units, the optimal structure will be decided in the process of learning. The objective of the paper is to explain the structural learning of both hierarchical and mutual connecting neural networks. Both networks obtained and showed the sufficiently good results. In the stock forecast by a general neural network, the operation and the system cost are very large because a lot of numbers of hidden layer units in the network are used. This research tried the optimization of the network by the structured learning, and evaluated the practicality.
UR - http://www.scopus.com/inward/record.url?scp=37249017033&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=37249017033&partnerID=8YFLogxK
M3 - Conference contribution
AN - SCOPUS:37249017033
SN - 9783540729174
VL - 4529 LNAI
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 231
EP - 240
BT - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
ER -