A weight-featured and data-distribution-based fuzzy pattern classification approach
Control and Intelligent Systems
| Hi-index | 0.00 |
This note analyzes the unsupervised fuzzy neural network (FNNU) of the original paper by Kwan and Cai (ibid., vol.2, p.185-93, 1994) and finds the following: the FNNU is a clustering net, not a classifier net, and the number of clusters the network settles to may be less or more than the actual number of pattern classes (sometimes it could even be equal to the number of training data points); the huge number of connections in the FNNU can be drastically reduced without degrading its performance; and the algorithm does not have any learning capability for its parameters. Computational experience shows that usually the performance of a multilayer perceptron (MLP) is comparable to that of even a supervised version of FNN (trained by gradient descent algorithm) in terms of recognition scores, but an MLP has a much faster convergence than the supervised version of FNN