Determining the Fault Tolerance of MemristorsBased Neural Network Using Simulation and Design of Experiments

The system approach to determining the fault tolerance of artificial neural networks (ANN) implementing with the use of a new electronic element named memristor is proposed in the article. The approach is invariant to an ANN architecture, a scheme of memristors-based components embodiment, a type of memristors and a task to be solved by ANN. The proposed approach is based on the application of the methodologies of simulation and designing of experiments. With the help of the method developed by the authors, the determination and investigation of fault tolerance of the memristors-based ANN (ANNM) which was trained to recognize of handwritten digits from the MNIST database were done. In this example, it is shown that the fault tolerance of ANNM depends not only on hardware or software embodiment tools but also on its information parameters, which are chosen at the design stage. The proposed approach and method are intended for application in the engineering design process of the ANNM and allows to provide the necessary level of their fault tolerance and dependability.