Development of an Algorithm for Detection and Recovery of Corruption in Convolutional Neural Networks Data Storage

Abstract

Computer vision based applications are commonly utilized in embedded systems. The demand for higher accuracy leads to increased complexity of convolutional neural networks (CNN) and hence, larger storage requirement for saving pre-trained networks. Even the smallest data corruption in the storage units of the embedded systems can result in drastic failures due to the propagation of the errors. This thesis proposes a new algorithm for the recovery of the data in case of single event upset (SEU) error. An association rule mining based algorithm is used to find the probability of corruption in each of the bits. The recovery algorithm is tested on four different trained ResNet and the best recovery rate of 66% was found in the most complex scenario. However, for the special cases of corruption in the frequently repeated bits, the recovery rate was found to be perfect with 100% recovery rate.