Selective Hardening of CNNs based on Layer Vulnerability Estimation

There is an increasing interest in employing Convolutional Neural Networks (CNNs) in safety-critical application fields. In such scenarios, it is vital to ensure that the application fulfills the reliability requirements expressed by customers and design standards. On the other hand, given the CNNs extremely high computational requirements, it is also paramount to achieve high performance. To meet both reliability and performance requirements, partial and selective replication of the layers of the CNN can be applied. In this paper, we identify the most critical layers of a CNN in terms of vulnerability to fault and selectively duplicate them to achieve a target reliability vs. execution time trade-off. To this end we perform a design space exploration to identify layers to be duplicated. Results on the application of the proposed approach to four case study CNNs are reported.