2019/09/30 by Christoph Schorn, Schorn, Christoph, Thomas Elsken +9
Computer Science · Engineering · #Advanced Memory and Neural Computing #Advanced Neural Network Applications #FOS: Computer and information sciences #Machine Learning (cs.LG) #Machine Learning (stat.ML) #Radiation Effects in Electronics
paper · pdf · doi:10.48550/arxiv.1909.13844
openalex publication_date 2019/09/30 · openalex created_date 2022/07/28 · openalex updated_date 2026/07/28
Applying deep neural networks (DNNs) in mobile and safety-critical systems,\nsuch as autonomous vehicles, demands a reliable and efficient execution on\nhardware. Optimized dedicated hardware accelerators are being developed to\nachieve this. However, the design of efficient and reliable hardware has become\nincreasingly difficult, due to the increased complexity of modern integrated\ncircuit technology and its sensitivity against hardware faults, such as random\nbit-flips. It is thus desirable to exploit optimization potential for error\nresilience and efficiency also at the algorithmic side, e.g., by optimizing the\narchitecture of the DNN. Since there are numerous design choices for the\narchitecture of DNNs, with partially opposing effects on the preferred\ncharacteristics (such as small error rates at low latency), multi-objective\noptimization strategies are necessary. In this paper, we develop an\nevolutionary optimization technique for the automated design of\nhardware-optimized DNN architectures. For this purpose, we derive a set of\neasily computable objective functions, which enable the fast evaluation of DNN\narchitectures with respect to their hardware efficiency and error resilience\nsolely based on the network topology. We observe a strong correlation between\npredicted error resilience and actual measurements obtained from fault\ninjection simulations. Furthermore, we analyze two different quantization\nschemes for efficient DNN computation and find significant differences\nregarding their effect on error resilience.\n