Evaluating Convolutional Neural Networks Reliability depending on their Data Representation

Abstract

Safety-critical applications are frequently based on deep learning algorithms. In particular, Convolutional Neural Networks (CNNs) are commonly deployed in autonomous driving applications to fulfil complex tasks such as object recognition and image classification. Ensuring the reliability of CNNs is thus becoming an urgent requirement since they constantly behave in human environments. A common and recent trend is to replace the full-precision CNNs to make way for more optimized models exploiting approximation paradigms such as reduced bit-width data type. If from one hand this is poised to become a sound solution for reducing the memory footprint as well as the computing requirements, it may negatively affect the CNNs resilience. The intent of this work is to assess the reliability of a CNN-based system when reduced bit-widths are used for the network parameters (i.e., synaptic weights). The approach evaluates the impact of permanent faults in CNNs by adopting several bit-width schemes and data types, i.e., floating-point and fixed-point. This determines the trade-off between the CNN accuracy and the bits required to represent network weights. The characterization is performed through a fault injection environment built on the darknet open source framework. Experimental results show the effects of permanent fault injections on the weights of LeNet-5 CNN.