Journal
NEURAL NETWORKS
Volume 94, Issue -, Pages 239-254Publisher
PERGAMON-ELSEVIER SCIENCE LTD
DOI: 10.1016/j.neunet.2017.07.004
Keywords
Fully tuned RBFNNs; Improved maximum data coverage algorithm; Specialized Orthant-Wise Limited-memory; Quasi-Newton method; L-1 regularization; Classification
Funding
- National Natural Science Foundation of China [61273122, 61005047]
- Qing Lan Project of Jiangsu Province
- NPRP from the Qatar National Research Fund (a member of Qatar Foundation) [NPRP8-274-2-107]
Ask authors/readers for more resources
This paper investigates the construction of sparse radial basis function neural networks (RBFNNs) for classification problems. An efficient two-phase construction algorithm (which is abbreviated as TPCLR1 for simplicity) is proposed by using L-1 regularization. In the first phase, an improved maximum data coverage (IMDC) algorithm is presented for the initialization of RBF centers and widths. Then a specialized OrthantWise Limited-memory Quasi-Newton (sOWL-QN) method is employed to perform simultaneous network pruning and parameter optimization in the second phase. The advantages of TPCLR1 lie in that better generalization performance is guaranteed with higher model sparsity, and the required storage space and testing time are much reduced. Besides these, only the regularization parameter and the maximum number of function evaluations are required to be prescribed, then the entire construction procedure becomes automatic. The learning algorithm is verified by several classification benchmarks with different levels of complexity. The experimental results show that an appropriate value of the regularization parameter is easy to find without using costly cross validation, and the proposed TPCLR1 offers an efficient procedure to construct sparse RBFNN classifiers with good generalization performance. (C) 2017 Elsevier Ltd. All rights reserved.
Authors
I am an author on this paper
Click your name to claim this paper and add it to your profile.
Reviews
Recommended
No Data Available