Efficient identification of critical faults in a memristor crossbar-based deep neural network to improve reliability for deep learning applications

Unmet Need

Deep neural networks have been employed in a wide range of deep learning applications ranging from autonomous vehicles and medical diagnosis to smart agriculture and image recognition. Specialized neuromorphic hardware, such as the memristor-based crossbars, provide an energy-efficient way to implement deep learning algorithms. However, due to various types of manufacturing defects and process variations, hardware-level reliability is a key concern for memristor crossbars. At the same time, these architectures are inherently fault-tolerant, and many faults do not have any significant impact on the accuracy of neural networks. Therefore, there is an urgent need to develop a framework to efficiently identify critical faults in the crossbar.


Researchers at Duke have invented a method to locate critical faults in memristor crossbar-based deep neural networks (DNN). This method can be used to improve the reliability of DNN hardware systems that are used for deep learning applications. Specifically, this is a misclassification-driven training algorithm to efficiently identify critical faults in the crossbar. This technology employs a trained machine learning (ML) model to predict whether a fault is catastrophic. The low-cost fault-tolerant framework can achieve a high degree of classification accuracy and targets only critical faults to reduce the overhead. Experimental results for the DNN architectures, AlexNet and VGG-16 on the CIFAR-10 data sets, showed that this method can accurately and rapidly identify a large number of CFs and has a remarkably good model transferability to different data sets, e.g., CIFAR-100 and ImageNet.


  • Targets only the critical faults to reduce the hardware and performance overhead for fault tolerance, and be efficiently used for deep learning applications
  • More than 50 times faster than a baseline method in identifying CFs
  • Accurately identifies 99% of CFs
  • Highly transferable among popular data sets
Central processing unit chip on circuit board

Duke File (IDF) Number



  • Chen, Ching-Yuan
  • Chakrabarty, Krishnendu

For more information please contact


Pratt School of Engineering