This paper considers an alternative activation function for use with MLP networks. The performance on parity problems is considered and it has been found that only n — 1 hidden units were needed to resolve the n-bit problem. Also, insight has been gained into the families of network parameters generated. Use as the kernel of a support vector machine for particular problems is anticipated.
KeywordsSupport Vector Machine Radial Basis Function Activation Function Sigmoid Function Radial Basis Function Network
Unable to display preview. Download preview PDF.
- Menoll A. Mehrotra K, Mohan C, Ranka S: Characterization of a Class of Sigmoid Functions with Applications to Neural Networks. Neural Networks 9, No 5, pp 818–835, (1996).Google Scholar
- Reeves C. R. Johnston C.: Fitting densities and hazard functions with neural networks. Accepted by ICANNGA 2001.Google Scholar