On the analysis of the criticalities of deep neural network.

In recent years, the complexity of emerging computing systems has led to the need of developing new algorithms and solutions for both industry and academy. In this scenario, due to their outstanding computational capabilities, artificial neural networks (ANNs) have become attractive solutions in tasks such as image classification, such as radars, flight control, robots, self-driving cars, space applications. However, to adopt this solution safely in human context, it is crucial to assess their reliability. This work presents a methodology to identify the most critical neurons of a convolutional neural networks (CNN) (Resnet-18) trained on two different datasets, i.e. MNIST and CIFAR-10. This method is based on two levels of analysis: first, the neuron is viewed as an element of each output class (class-oriented analysis); second, the same is interpreted as belonging to the entire neural network (network-oriented analysis). In detail, a neuron that contributes more to the final prediction of the network and therefore carries more information than others, was considered critical. The proposed methodology has been validated by means of two software fault injection campaigns, through which we confirmed that the accuracy of the network decreased much faster in the case of fault involving neurons defined as critical with respect to neurons of minor importance within the network. Subsequently, to improve the reliability of the target convolutional neural networks, a strategy based on the Triple Modular Redundancy (TMR) technique was adopted. The goal was to identify a mitigation technique able to protect all the neurons labeled as critical in the previous step in case they were subject to a fault. The TMR is a fault masking technique that, in our study, allowed to triple all those critical neurons and, through the usage of a majority voter, propagate in the successive layers of the network the most common output of the said neuron, masking possible single faults. In particular, through software fault injection campaigns on different percentages of neurons , we have applied the TMR technique on all the neurons deemed critical and we have shut off all the outputs of those neurons considered not important for the elaboration of the final forecast by the network. Using this method of shutdown parallel to the tripling, we have investigated the trade-off between memory footprint and reliability.