Discrete neural networks and fingerprint identification

Abstract

The author has developed a general method for discretization of feedforward neural networks and has empirically demonstrated the usefulness of the method by successfully applying it to the nontrivial task of fingerprint identification. Surprisingly, the discrete neural network (DNN) developed in this way demanded just 4 b for the table representation of the sigmoid function, and only 6 b for the representation of the matching discrete solution. It is clearly shown that there is no significant difference in the performance on the test set between the real neural network and the DNN. Thus, it is concluded that the discretization methods proposed have shown themselves to be realistic.<>