Defect-Tolerant Crossbar Training of Memristor Ternary Neural Networks

Abstract

A memristor Ternary Neural Network (TNN) is a promising candidate for implementing neural networks for Internet-of-Things (IoT) applications, where low power and simple hardware are important. One important concern in the memristor TNN is that the real memristor crossbar has various defects such as stuck-at-faults, memristance variation, etc., like human brain's biological neurons and synapses. To mitigate the inference loss due to the memristive defects, we need to retrain the defective crossbar. However, the crossbar's retraining needs a long time and a large amount of energy of programming, because the memristors should be programmed one by one using Incremental Step Pulse Programming (ISPP) of flash memories. Here, we combine the partial-gated training scheme with the asymmetrical training for not only minimizing the recognition rate loss, but also saving the crossbar's programming time and energy. The CF scheme with 10% retraining indicates the programming energy can be saved by as large as ∼98%, in sacrifice with the MNIST rate loss of ∼0.6%, compared to the FF scheme with 100% retraining. The simulation indicates that the CF with 10% retraining will be useful for realizing the crossbar-based neural-network in large scale for future IoT applications.