Learning method by overload

In this article, a new learning method to adjust the number of dimensions of patterns on hidden layers to the complexity of given tasks is presented. In this method, in addition to a given task, a controllable additional task is given to a network to adjust the size of the network and the capacity of tasks that the network learns. We found that this method reduces the number of dimensions of patterns for original tasks, and also the volume of patterns for the original task. This makes it easy to analyze hidden patterns symbolically and provides a generalization power to the networks.