Deep learning (DL) reliability is becoming a growing concern, and efficient reliability assessment approaches are required to meet safety constraints. This article presents a survey of the main DL reliability assessment methodologies, focusing mainly on fault injection techniques used to evaluate DL resilience.

Convolutional Neural Networks (CNNs) are currently one of the most widely used predictive models in machine learning. Recent studies have demonstrated that hardware faults induced by radiation fields, including cosmic rays, may significantly impact the CNN inference leading to wrong predictions. Therefore, ensuring the reliability of CNNs is crucial, especially for safety-critical systems. In the literature, several works propose reliability assessments of CNNs mainly based on statistically injected faults. This work presents a software emulator capable of injecting real faults retrieved from radiation tests. Specifically, from the device characterisation of a DRAM memory, we extracted event rates and fault models. The software emulator can reproduce their incidence and access their effect on CNN applications with a reliability assessment precision close to the physical one. Radiation-based physical injections and emulator-based injections are performed on three CNNs (LeNet-5) exploiting different data representations. Their outcomes are compared, and the software results evidence that the emulator is able to reproduce the faulty behaviours observed during the radiation tests for the targeted CNNs. This approach leads to a more concise use of radiation experiments since the extracted fault models can be reused to explore different scenarios (e.g., impact on a different application).

Nowadays, the usage of electronic devices running artificial neural networks (ANNs)-based applications is spreading in our everyday life. Due to their outstanding computational capabilities, ANNs have become appealing solutions for safety-critical systems as well. Frequently, they are considered intrinsically robust and fault tolerant for being brain-inspired and redundant computing models. However, when ANNs are deployed on resource-constrained hardware devices, single physical faults may compromise the activity of multiple neurons. Therefore, it is crucial to assess the reliability of the entire neural computing system, including both the software and the hardware components. This article systematically addresses reliability concerns for ANNs running on multiprocessor system-on-a-chips (MPSoCs). It presents a methodology to assign resilience scores to individual neurons and, based on that, schedule the workload of an ANN on the target MPSoC so that critical neurons are neatly distributed among the available processing elements. This reliability-oriented methodology exploits an integer linear programming solver to find the optimal solution. Experimental results are given for three different convolutional neural networks trained on MNIST, SVHN, and CIFAR-10. We carried out a comprehensive assessment on an open-source artificial intelligence-based RISC-V MPSoC. The results show the reliability improvements of the proposed methodology against the traditional scheduling.

The complexity of the state-of-the-art devices makes reliability assessments approaches extremely complex and, sometimes, out of the timing constraints and computational capabilities. Fault Injections (FIs) are one of the most used approaches for evaluating the dependability of safety-critical systems. With billion-transistor hardware devices running trillion-parameter deep neural networks, injecting the entire fault universe is unfeasible. A widespread solution consists in performing statistical fault injections (SFIs), injecting a subset of faults to estimate a characteristic with an error margin and a confidence level. This research work presents an iterative SFI approach to estimate failure rates in convolutional neural networks (CNNs), i.e., the percentage of wrong predictions caused by random hardware faults affecting synaptic weights. SFIs at different granularities have been performed with margin of errors equal to 1%, 0.1%, and 0.01%. Results for two CNNs (ResNet20 and MobileNetV2) are presented and experimentally and statistically demonstrate the effectiveness of the proposed approach. For instance, to estimate the network-wise failure rate with an error margin of 0.01%, the proposed approach reduces the total injected faults by about 66% and 90% compared to conservative methods, and by 1.94% and 1.65% compared to iterative SFI methods in the literature, for ResNet20 and MobileNetV2, respectively.