An effective test method for digital neural networks
Érika Cota, Luigi Carro, Marcelo Soares Lubaszewski · 2002
This paper presents some results of a new test methodology for neural networks, based on the mathematical similarities of these systems with digital filters. The method consists on defining the signature of the net by means of its own weights. During test, the network is reconfigured so that the weights are driven to the network inputs, and results of the net outputs are compared with the previously defined signature. The first experiments and results of this technique applied to a feedforward neural network for character recognition are presented. The proposed at-speed method is very effective for a processor-like network implementation. It requires low area overhead, while just small modifications in the network design are necessary. Moreover, the testing time varies linearly with the number of net weights and neurons.