This paper discusses the problem on construction of classification hyperplane in hidden space for feedforward neural networks.This problem directly corresponds to the design of the connection weights from hidden to output neurons.The research result shows that the connection weights from hidden neurons to the output neuron can all be +1

and the bias of the output neuron can take only a basic bias of 12 plus an integer in a limited integer range as an auxiliary bias.This results that the hidden to output connection weights and bias do not need to train

which decreases the searching space of network parameters during the training process of the networks

and it is much easier for the hardware realization of the network as there is no need for multiplicator

and the network constructed in this way is somewhat robust to the input data.