Approximate Down-Sampling Strategy for Power-Constrained Intelligent Systems

In modern power constrained applications, as with most of those belonging to the Internet-of-Things world, custom hardware supports are ever more commonly adopted deploy artificial intelligence algorithms. these operating environments, limiting dissipation much possible is mandatory, even at expense reduced computational accuracy. this paper we propose a novel prediction method identify potential predominant features in convolutional layers followed by down-sampling layers, thus reducing overall number convolution calculations. This approximation strategy has been exploited design architecture for inference Convolutional Neural Network (CNN) models. The proposed approach applied several benchmark CNN models and achieved an energy saving up 70% accuracy loss lower than 3%, respect baseline designs. Performed experiments demonstrate that, when infer Visual Geometry Group-16 (VGG16) network model, implemented on Xilinx Z-7045 chip STM 28nm process technology dissipates only 680 21.9 mJ/frame, respectively. both cases, overcomes state-of-the-art competitors terms energy-accuracy drop product.