Learning in the hypercube

A discussion on learning from a geometric point of view is presented. Its main purpose is to obtain ways of pre-estimating the weights and thresholds of a neural network and of analyzing the corresponding effects on the learning procedure speed. It is conceivable that even a rough preliminary evaluation of the position of the hyperplanes will improve the learning process. It is also conceivable that, in spaces with well-defined geometric properties, the information inherently contained in the training data set can be helpful in making such an evaluation. This subject is discussed, and basic alternatives for doing so are shown. It is concluded that, depending on the configuration of the space, geometric tools can indeed lead to improvements