【摘 要】
：
The RRAM-based accelerators enable fast and energy-efficient inference for neural networks.However,there are some requirements to deploy neural networks on RRAM-based accelerators,which are not considered in existing neural networks.(1) Because the noise
【机 构】
：
Center for Energy-Efficient Computing and Applications,Peking University,Beijing 100871,China;Depart
论文部分内容阅读
The RRAM-based accelerators enable fast and energy-efficient inference for neural networks.However,there are some requirements to deploy neural networks on RRAM-based accelerators,which are not considered in existing neural networks.(1) Because the noise problem and analog-digital converters/digital-analog converters (ADC/DAC) affect the prediction accuracy,they should be modeled in networks.(2) Be-cause the weights are mapped to the RRAM cells,they should be quantized,and the number of weights is limited by the number of RRAM cells in the accelerator.These requirements motivate us to customize the hardware-friendly network for the RRAM-based accelerator.We take the idea of network architecture search (NAS) to design networks with high prediction accuracy that meet the requirements.We propose a framework called NAS4RRAM to search for the optimal network on the given RRAM-based accelerator.The experiments demonstrate that NAS4RRAM can apply to different RRAM-based accelerators with different scales.The performance of searched networks outperforms the manually designed ResNet.
其他文献
The Internet of Things has created an increasing demand for giant magnetoresistive (GMR)sensor owing to its high sensitivity,low power-consumption and small size.A full Wheatstone bridge GMR sensor is fabricated on 6-inch wafers with an annealing process
Dear editor,rnFault-tolerant tracking control has recently become an area of research interest,in which fault detection,fault-tolerant control (FTC),and tracking-control techniques are inte-grated[1,2].It is noteworthy that disturbance is a criti-cal fact
Graph convolutional networks (GCNs) have been applied successfully in social networks and recommendation systems to analyze graph data.Unlike conventional neural networks,GCNs introduce an aggregation phase,which is both computation-and memory-intensive.T
Dear editor,rnMany complicated real-world systems can be considered as complex networks to analyze their macroscopic properties,e.g.,the Internet,power grids,and communication net-works[1].Recently,studies have found that many real-world networks are coup
Dear editor,rnThe fractional Fourier transform (FRFT) was proposed by Wiener in 1929.As an important and powerful analyzing tool for time-frequency analysis,the FRFT has been ap-plied in a lot of fields[1,2]such as signal processing,op-tics,radar,and quan
Dear editor,rnThe study of nonlinear time-delay systems made great progress in the past few decades[1]and some effective meth-ods were presented for special systems such as the approxi-mate linear method,sum of squares decomposition method,and nonlinear m
Dear editor,rnIn recent years,we have witnessed a rapid development of the side-channel attacks,which deviate from the tradi-tional block-box model and exploit the leakages of crypto-graphic devices.Among those attacks,the so-called differen-tial power an
Dear editor,rnOrthogonal frequency division multiplexing (OFDM) has been adopted in new radio (NR)[1],and multiple-port de-modulation reference signal (DMRS) has been designed to achieve channel estimation.In NR,two ports of DMRS share the same frequency
Dear editor,rnTo maintain the confidentiality of the sensitive data,users tend to encrypt their data under an associated access pol-icy (or attributes) beforeoutsourcing them to the cloud.In the traditional access control model via cryptographic approache
Privacy issues in control systems are getting more and more attention from both industry and research.In recent years,information technology and artificial intelligence technol-ogy are being increasingly employed in emerging applica-tions such as the Inte