Binary to binary mapping for classification plays an important role in the researches on feed-forward-neural-network learning．In this paper，the geometrical method is employed to work out a new algorithm to train binary neural networks for classification．By analysis of every training vertex's geometrical location，the algorithm alwavs produces a neural network of four layers for a certain classification problem．The advantages of this algorithm are：it runs with guaranteed convergence and goes to converge much more quickly than BP and some other algorithms；it can determine the structure of the neural networks by learning SO that a precise classification is carried out．In addition，every neuron generated by the algorithm employs a hard-limit activation func-tion with integer synaptic weights，which makes the actual implementation by VLSI tech-nology more facilitated．