Testing recurrent artificial neural networks

This paper presents a new and novel testing approach for detecting interconnection deletion faults in hardware implementations of artificial neural networks (ANNs). The proposed testing approach is based on an unusual ANN behavior manifested by faulted ANNs having apparent better performance than fault-free ANNs when neurons are operated with low activation function gains. Although transient, this noncoherent behavior can be used to detect interconnection deletion faults in ANNs. Using mathematical and simulation models, the efficacy of the low activation gain fault detection (LAGFD) method for fault detection is determined. Further, suggested design rules are stated enabling reasonable detection of interconnection deletion faults. A test algorithm is described that identifies a fault-free subset of an ANN using LAGFD. Finally, simulation examples are presented to verify the LAGFD test algorithm