Can anyone tell me the advantages of neural network classifiers? Neural network with Softmax transformation, or probabilistic neural network, is used in many classification system. The basic idea is to train the network to approximate underlying conditional probability. Many training algorithms use KL difference between the true probability and the probability represented by the network. However, the true probability is unknown, so the frequency is used in KL difference. So, when network training finished, the probability represented by the network would be close to the frequency. Why not use frequency directly? What is the advantage of Neural network classifier?
