FEEDFORWARD NEURAL NETWOTKS WITHOUT ORTHONORMALIZATION

Lei Chen, Hung Keng Pung, Fei Long

2007

Abstract

Feedforward neural networks have attracted considerable attention in many fields mainly due to their approximation capability. Recently, an effective noniterative technique has been proposed by Kaminski and Strumillo(Kaminski and Strumillo, 1997), where kernel hidden neurons are transformed into an orthonormal set of neurons by using Gram-Schmidt orthonormalization. After this transformation, neural networks do not need recomputation of network weights already calculated, therefore the orthonormal neural networks can reduce computing time. In this paper, we will show that it is equivalent between neural networks without orthonormal transformation and the orthonormal neural networks, thus we can naturally conclude that such orthonormalization transformation is not necessary for neural networks. Moreover, we will extend such orthonormal neural networks into additive neurons. The experimental results based on some benchmark regression applications further verify our conclusion.

References

  1. Huang, G.-B., Zhu, Q.-Y., and Siew, C.-K. (2006). Extreme learning machine: Theorey and applications. Neurocomputing, 70:489-501.
  2. Kaminski, W. and Strumillo, P. (1997). Kernel orthonormalization in radial basis function neural networks. IEEE Transactions On Neural Networks, 8(5):1177-1183.
Download


Paper Citation


in Harvard Style

Chen L., Keng Pung H. and Long F. (2007). FEEDFORWARD NEURAL NETWOTKS WITHOUT ORTHONORMALIZATION . In Proceedings of the Ninth International Conference on Enterprise Information Systems - Volume 2: ICEIS, ISBN 978-972-8865-89-4, pages 420-423. DOI: 10.5220/0002374704200423


in Bibtex Style

@conference{iceis07,
author={Lei Chen and Hung Keng Pung and Fei Long},
title={FEEDFORWARD NEURAL NETWOTKS WITHOUT ORTHONORMALIZATION},
booktitle={Proceedings of the Ninth International Conference on Enterprise Information Systems - Volume 2: ICEIS,},
year={2007},
pages={420-423},
publisher={SciTePress},
organization={INSTICC},
doi={10.5220/0002374704200423},
isbn={978-972-8865-89-4},
}


in EndNote Style

TY - CONF
JO - Proceedings of the Ninth International Conference on Enterprise Information Systems - Volume 2: ICEIS,
TI - FEEDFORWARD NEURAL NETWOTKS WITHOUT ORTHONORMALIZATION
SN - 978-972-8865-89-4
AU - Chen L.
AU - Keng Pung H.
AU - Long F.
PY - 2007
SP - 420
EP - 423
DO - 10.5220/0002374704200423