E2CNNs: Ensembles of Convolutional Neural Networks to Improve Robustness Against Memory Errors in Edge-Computing Devices

To reduce energy consumption, it is possible to operate embedded systems at sub-nominal conditions that introduce errors in their memories. These errors affect the CNN weights and activations, compromising accuracy. In this paper, we introduce Embedded Ensemble CNNs (E2CNNs), our design methodology to conceive ensembles of CNNs improving robustness against memory errors compared to a single-instance network. Unfortunately, SoA ensembles do not suit well embedded systems, in which memory constraints limit the number of deployable models. Our proposed architecture solves that limitation applying SoA compression methods to produce an ensemble with the same memory requirements of the original architecture, but with improved error robustness. Then, as part of our new E2CNNs design methodology, we propose a heuristic to automate the design of the ensemble that maximizes accuracy for an expected memory error rate while bounding the required design effort. To evaluate the robustness of E2CNNs for different error types, we propose three error models that simulate the behavior of SRAM and eDRAM at sub-nominal conditions. We show that E2CNNs achieves energy savings of up to 80% for LeNet-5, 90% for AlexNet, 60% for GoogLeNet, 60% for MobileNet and 60% for an industrial CNN, while minimizing the impact on accuracy.