A new method for inverting feedforward neural networks
Y. Araki, T. Ohki, Daniel Citterio, M. Hagiwara, Kunio Suzuki · 2004
In this paper, we propose a new method for inverting feedforward neural networks. Inversion of neural networks means to find the inputs which produce given outputs. In general, this is an ill-posed problem whose solution isn't unique. Inversion using iterative optimization method (for example gradient descent, quasi-Newton method) is useful to this problem and it is called "iterative inversion". We propose a new iterative inversion using a Bottleneck Neural Network with Hidden layer's input units (BNNH), which we design on the basis of Bottleneck Neural Network (BNN). Compressing input space by BNNH, we reduce the dimension of search space, or input space to be searched with iterative inversion. With reduction of the search space's dimension, performance about computation time and accuracy is expected to become better. In experiments, the proposed method is applied to some examples. These results show the effectivity of the proposed method.