Dynamic tunneling based regularization in feedforward neural networks

摘要

This paper presents a new regularization method based on dynamic tunneling for enhancing generalization capability of multilayered neural networks. The proposed method enables escape through undesired sub-optimal solutions on the composite error surface by means of dynamic tunneling. Undesired sub-optimal solutions may be increased or introduced from regularized objective function. Hence, the proposed method is capable of enhancing the regularization property without getting stuck at sub-optimal values in search space. The regularization property and escape from the sub-optimal values have been demonstrated through computer simulations on two examples.