UM  > 科技學院  > 電腦及資訊科學系
A new dynamic optimal learning rate for a two-layer neural network
Zhang T.1; Chen C.L.P.1; Wang C.-H.2; Tam S.C.1
Source PublicationProceedings 2012 International Conference on System Science and Engineering, ICSSE 2012
AbstractThe learning rate is crucial for the training process of a two-layer neural network (NN). Therefore, many researches have been done to find the optimal learning rate so that maximum error reduction can be achieved in all iterations. However, in this paper, we found that the best learning rate can be further improved. In saying so, we have revised the direction to search for a new dynamic optimal learning rate, which can have a better convergence in less iteration count than previous approach. There exists a ratio k between out new optimal learning rate and the previous one after the first iteration. In contrast to earlier approaches, the new optimal learning rate of the two-layer NN has a better performance in the same experiment. So we can conclude that our new dynamic optimal learning rate can be a very useful one for the applications of neural networks. © 2012 IEEE.
Keywordlearning rate neural network new optimal learning rate ratio k two-layer NN
URLView the original
Fulltext Access
Citation statistics
Document TypeConference paper
Affiliation1.Universidade de Macau
2.National Chiao Tung University Taiwan
Recommended Citation
GB/T 7714
Zhang T.,Chen C.L.P.,Wang C.-H.,et al. A new dynamic optimal learning rate for a two-layer neural network[C],2012:55-59.
Related Services
Recommend this item
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[Zhang T.]'s Articles
[Chen C.L.P.]'s Articles
[Wang C.-H.]'s Articles
Baidu academic
Similar articles in Baidu academic
[Zhang T.]'s Articles
[Chen C.L.P.]'s Articles
[Wang C.-H.]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[Zhang T.]'s Articles
[Chen C.L.P.]'s Articles
[Wang C.-H.]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.