A hybrid multilayer neural network for binary pattern classification and its low-bit learning algorithm

A hybrid multilayer neural network and its low-bit learning algorithm for binary pattern classification are proposed. In the training process, a single-layer network is first employed. If the training does not converge, then a middle unit is assigned to a critical pattern. Connection weights are adjusted so that this unit responds to the critical pattern. The training is repeated by increasing the middle units. Thus, the number of middle units can be optimized. Connection weights are adjusted using a small number of bits, resulting in very simplified digital hardware. Since the outputs of the middle and the output layers are specified, a single-layer low-bit learning algorithm is proposed. The training is very fast and is insensitive to initial weights and parameters. A divided form is proposed in order to drastically save the middle units for a large number of the patterns