The special character of certain degrees of freedom in two-layered neural networks is investigated for on-line learning of realizable rules. Our analysis shows that the dynamics of these degrees of freedom can be put on a faster timescale than those remaining, with the profit of speeding up the overall adaptation process. This is shown for two groups of degrees of freedom: second-layer weights and bias weights. For the former case our analysis provides a theoretical explanation of phenomenological findings. The resulting learning algorithm is compared with natural gradient descent in order to check whether the proposed scaling can be naturally derived from that type of learning rule.