Some properties of multilayer neural networks with different learning coefficients for each layer

In this paper, we discuss some properties of multilayer neural networks with different learning coefficients for each layer, using different learning backpropagation (DLBP). By changing the ratio of the learning coefficients, we can control learning progress at each layer. And we can also control the influence of each hidden units on the network. With DLBP learning, we can construct the corresponding networks to the opposite requirements such as to set the network with the minimum hidden units by eliminating the hidden units which have little influence, or to get fault tolerant network by balancing the influence of all hidden units. In DLBP learning, the evaluation function is not changed, and convergence to the error minimum solution is guaranteed.