Abstract
Applying deep neural networks (DNNs) in mobile and safety-critical systems,such as autonomous vehicles, demands a reliable and efficient execution onhardware. Optimized dedicated hardware accelerators are being developed toachieve this. However, the design of efficient and reliable hardware has becomeincreasingly difficult, due to the increased complexity of modern integratedcircuit technology and its sensitivity against hardware faults, such as randombit-flips. It is thus desirable to exploit optimization potential for errorresilience and efficiency also at the algorithmic side, e.g., by optimizing thearchitecture of the DNN. Since there are numerous design choices for thearchitecture of DNNs, with partially opposing effects on the preferredcharacteristics (such as small error rates at low latency), multi-objectiveoptimization strategies are necessary. In this paper, we develop anevolutionary optimization technique for the automated design ofhardware-optimized DNN architectures. For this purpose, we derive a set ofeasily computable objective functions, which enable the fast evaluation of DNNarchitectures with respect to their hardware efficiency and error resiliencesolely based on the network topology. We observe a strong correlation betweenpredicted error resilience and actual measurements obtained from faultinjection simulations. Furthermore, we analyze two different quantizationschemes for efficient DNN computation and find significant differencesregarding their effect on error resilience.