Vulnerability of Hardware Neural Networks to Dynamic Operation Point Variations
Jeng-Hau Lin, Xun Jiao, Mulong Luo, Zhuowen Tu, Rajesh K. Gupta · IEEE Design and Test · 2020
Editor's notes: This article studies the impacts of physical variations on neural networks. The proposed studies reveal an important observation that both multiple layer perceptron (MLP) and convolutional neural network (CNN) may fail to operate appropriately even with small variations (e.g., voltage droops as small as 20 mV). Robust neural network architectures, including the binarized neural network (BNN) and the local binary pattern network (LBPNet), are explored to address this variability issue that has become a major bottleneck for practical applications. -Xin Li, Duke University.