Strategies for Improving the Error Robustness of Convolutional Neural Networks

Abstract

The error robustness of Convolutional Neural Networks (CNNs) is an important attribute requiring attention due to their growing application in safety-critical domains such as autonomous driving and medical devices. Hardware errors affecting the execution of such models may lead to system failures and, therefore, fault tolerance techniques are necessary to improve dependability. This paper proposes an approach to improve the robustness of CNNs and experimentally compares it with three other existing techniques. Fault injection is used to emulate hardware faults affecting CNNs targeting four distinct datasets. Results indicate that the ranger technique globally provides the best robustness closely followed by the stimulated training technique, although the former provides much lower temporal overhead than the latter. Architectural redundancy and dropout provide varying results. In all cases, caution through final evaluation of any CNN is required, because there are corner cases in which the robustness decreases, contrary to the intended outcome.