A digital multilayer neural network with limited binary expressions

A design methodology is presented for digital multilayer neural networks with limited binary expressions. An error back-propagation algorithm is modified as follows: the number of binary bits used for connections and unit outputs are decreased step by step in the training process. In order to express unit outputs with two-level values, the differential of the logistic function is replaced by a small positive constant used in weight change equations. After the training is completed, binary expressions for connections and unit outputs can be reduced to several-bit and two-level values, respectively. Therefore, no multipliers or nonlinear functions are required in the resulting network, which will be used for pattern recognition. Furthermore, memory capacity and adder circuit hardware can be reduced. The network performance is also insensitive to noisy patterns