Tien Nguyen Van
Ternary neural network with three values of the weight (-1, 0, +1) is an effective solution for artificial intelligence applications in edge computing due to their energy efficiency, which is achieved through their deployment on memristor crossbars. However, the implementation on a memristor crossbar often has problems related to programming energy and defects. Therefore, this paper proposes a training method that uses both weight pruning and defect-aware training to optimize the number of programming pulses while maintaining a high recognition rate for the network. By using this method, at a 20% defect rate and 80% sparsity, the network achieves a 95% recognition rate on MNIST and 90% on CIFAR-10, compared to only 73.7% and 60.4% without using defect-aware training. At the same sparsity, the number of programming pulses, which is indicative of programming energy, is reduced by 49.1% compared to 20% sparsity