A new method for inverting nonlinear multilayer feedforward networks

A novel method for inverting a mapping of the multilayer feedforward network is proposed. This method is based upon recursive constrained linear equations constructed by a given desired output, weights, an activation function and unknown inputs. By solving these recursive constrained linear equations by nonlinear and linear programming techniques, some typical inversions from a given output can be derived. Therefore, some new relations can be obtained between the unseen inputs and the given output. Two algorithms based on this method for inverting the mapping of the three-layer feedforward network are derived. The first algorithm is to invert the mapping into a global input space. The second algorithm is developed to invert the mapping into a local input space