DocumentCode :
3496576
Title :
Training neural network with damped oscillation and maximized gradient function
Author :
Islam, Mobarakol ; Khan, Md Tofael Hossain ; Rahaman, Arifur ; Saha, Samar K. ; Kundu, Anindya Kumar ; Rana, Md Masud
Author_Institution :
Dept. of Electron. & Commun. Eng., Khulna Univ. of Eng. & Technol., Khulna, Bangladesh
fYear :
2011
fDate :
22-24 Dec. 2011
Firstpage :
532
Lastpage :
537
Abstract :
Constant learning rate (LR) which is most widely used for training neural networks (NNs) in back propagation (BP) but it is not usually preferable due to its slow convergence rate while using small learning rate and it also shows less accuracy while using higher learning rate. In this paper, we are proposing a faster and supervised algorithm which shows more accuracy in a few iterations while dealing with neural networks (NNs). Training of NNs with damped oscillation and maximized gradient function (DOMG) deals with the implementation of damped oscillation in learning rate called damped learning rate (DLR) by which we get more accuracy in a few iterations and maximized gradient function is used for fast weight updating. DOMG is significantly tested on eight real world benchmark classification problems such as heart disease, ionosphere, Australian credit card, time series, wine, horse, glass and soybean identification. The proposed DOMG outperforms the existing BP in terms of convergence rate and generalization ability.
Keywords :
backpropagation; gradient methods; neural nets; pattern classification; Australian credit card; back propagation; benchmark classification problems; constant learning rate; convergence rate; damped learning rate; damped oscillation and maximized gradient function; fast weight updating; generalization ability; glass; heart disease; horse; ionosphere; neural network training; soybean identification; supervised algorithm; time series; wine; Artificial neural networks; Classification algorithms; Glass; Convergence rate; Damped learning rate; Maximized gradient function; Neural Networks;
fLanguage :
English
Publisher :
ieee
Conference_Titel :
Computer and Information Technology (ICCIT), 2011 14th International Conference on
Conference_Location :
Dhaka
Print_ISBN :
978-1-61284-907-2
Type :
conf
DOI :
10.1109/ICCITechn.2011.6164846
Filename :
6164846
Link To Document :
بازگشت