Association Rule Mining Based Algorithm for Recovery of Silent Data Corruption in Convolutional Neural Network Data Storage

Abstract

Embedded systems are finding their way into almost every aspects of our daily life from mp3 players and console games to the mobile phones. Different Artificial Intelligence (AI) based applications are commonly utilized in embedded systems from which computer vision based approaches are included. The demand for higher accuracy in computer vision applications is associated with the increased complexity of convolutional neural networks and the storage requirement for saving pre-trained networks. Different factors can lead to the data corruption in the storage units of the embedded systems, which can result in drastic failures due to the propagation of the errors. Hence, the development of software-based algorithms for the detection and recovery of data corruption is crucial for improvement and failure-prevention of embedded systems. This paper proposes a new algorithm for the recovery of the data in the case of single event upset (SEU) error. The association rule mining based algorithm will be used to find the probability of the corruption in each of the bits. The recovery algorithm was tested on four different pre-trained ResNet (ResNet32 and ResNet110 at two different accuracy levels each) and the best recovery rate of 66% was found in the most complex scenario, i.e., random bit corruption. However, for the special cases of SEU errors, e.g. error in the frequently repeated bits, the recovery rate was found to be perfect with a value of 100%.