Fuzzy signature neural networks for classification: Optimising the structure

We construct fuzzy signature neural networks where fuzzy signatures replace hidden neurons in a neural network similar to a radial basis function neural network. We investigated the properties of a naïve and a principled approach to fuzzy signature construction. The naïve approach provides very good results on benchmark datasets, but is outperformed by the principled approach when we approximate the noisy nature of real world datasets by randomly eliminating 20% of the data. The major benefit of the principled approach is to substantially improve robustness of the fuzzy signature neural networks we produce.