In this paper, we present a new procedure which relates the convergence rate of the traditional back-propagation algorithm with the sequence in which we must present the exemplar patterns to the network. According to this approach, we present the exemplar patterns to the network in a sequence which depends on the learning "difficulty" that each exemplar pattern imposes to the network. So, although all the exemplar patterns are learned, some of them are exposed to the network more frequently than others and the frequency rate is proportional to the learning "difficulty" that each exemplar pattern imposes. We have tested this procedure with a representative class of problems. This procedure decreases dramatically the number of learning sweeps without increasing the overall computational complexity of the algorithm. Thus, the timing requirements of the learning phase are greatly reduced without sacrificing the precision in the retrieving phase.