Gradient descent training of neural networks can be done in either a batch or on-line manner. A widely held myth in the neural network community is that batch training is as fast or faster and/or more ‘correct’ than on-line training because it supposedly uses a better approximation of the true gradient for its weight updates. This paper explains why batch training is almost always slower than on-line training—often orders of magnitude slower—especially on large training sets. The main reason is due to the ability of on-line training to follow curves in the error surface throughout each epoch, which allows it to safely use a larger learning rate and thus converge with less iterations through the training data. Empirical results on a large (20,000-instance) speech recognition task and on 26 other learning tasks demonstrate that convergence can be reached significantly faster using on-line training than batch training, with no apparent difference in accuracy.