Discrimination ability of neural network and maximum likelihood classifiers

Artificial neural networks have been taken as powerful tools for pattern recognition and data analysis. However, the previous classification results show differences between the neural network classifier (NNC) and the statistical methods, such as maximum likelihood classifier (MLC). The reasons for the differences in classification accuracy were not completely understood. Much research has explored the behaviour of NNCs, including inputs, number of hidden layer(s), number of nodes, size of sample sets, training parameters. However, little work has been done on the effect of the overlap degree of patterns in feature space on a NNC. This study analyses the effect of the overlap degree of classes in feature space on performance of a standard backpropagation NNC and compares the results with MLC. Two data sets (i.e., the simulated data sets with different overlap degrees and the remotely sensed imagery with more complicate overlap pattern) are used to test the performance of a neural network. One surprising result is that NNC failed to discriminate two non-overlapping classes. Another interesting result is that NNC classifies class2 better while MLC classifies class1 better. "Z" statistics is carried out to compare two classifiers and results show that NNC is significantly better than MLC in the case of using simulated data at 95% C.I., but not significantly different with MLC in the case of using actual image in the study.