Memristive Artificial Neural Networks Accuracy Evaluation Using Data-Driven Models

Memristive devices are promising candidates for the role of analog synapses in the hardware implementation of energy-efficient artificial neural networks (ANNs). However, variations in the resistance of memristive devices lead to errors in the values of the weighting coefficients of neuron synapses and, accordingly, can have a negative impact on the accuracy of the functioning of ANNs based on them. This paper proposes an approach to creating models of memristive devices that can be used to assess the impact of their variations on the quality of ANN functioning. The implementation of this approach is demonstrated using the example of an ANN trained to solve a image classification problem.