Fully hardware-implemented memristor convolutional neural network

Yao, Peng; Wu, Huaqiang; Gao, Bin; Tang, Jianshi; Zhang, Qingtian; Zhang, Wenqiang; Yang, J. Joshua; Qian, He

doi:10.1038/s41586-020-1942-4

Cited by 1,560 publications

(1,150 citation statements)

References 33 publications

Supporting

Mentioning

1,138

Contrasting

Unclassified

Order By: Relevance

“…[ 1 ] In memristive neural networks, synapses and neuronal units are essential elements. Thanks to the compact structure and intrinsic learning ability, [ 2‐8 ] and the advantages of parallel computing and low energy consumption, [ 2,9‐15 ] memristive devices are commonly used as artificial synapses. However, the conductance (weight) of many memristive devices is modulated nonlinearly and asymmetrically during training, [ 16‐21 ] which limits the classification accuracy.…”

Section: Figurementioning

confidence: 99%

Implementation of Dropout Neuronal Units Based on Stochastic Memristive Devices in Neural Networks with High Classification Accuracy

Huang

Yang

et al. 2020

Advanced Science

View full text Add to dashboard Cite

Neural networks based on memristive devices have achieved great progress recently. However, memristive synapses with nonlinearity and asymmetry seriously limit the classification accuracy. Moreover, insufficient number of training samples in many cases also have negative effect on the classification accuracy of neural networks due to overfitting. In this work, dropout neuronal units are developed based on stochastic volatile memristive devices of Ag/Ta2O5:Ag/Pt. The memristive neural network using the dropout neuronal units effectively solves the problem of overfitting and mitigates the negative effects of the nonideality of memristive synapses, eventually achieves a classification accuracy comparable to the theoretical limit. The stochastic and volatile switching performances of the Ag/Ta2O5:Ag/Pt device are attributed to the stochastical rupture of the Ag filament under high electrical stress in the Ta2O5 layer, according to the TEM observation and the kinetic Monte Carlo simulation.

show abstract

Section: Figurementioning

confidence: 99%

Implementation of Dropout Neuronal Units Based on Stochastic Memristive Devices in Neural Networks with High Classification Accuracy

Huang

Yang

et al. 2020

Advanced Science

View full text Add to dashboard Cite

show abstract

“…For potential application in neuromorphic computing, memristors or other elements are often integrated in statistic fiber networks [56][57][58], an approach which cannot be realized with a single-step electrospinning process. Here, instead, we report on a combination of magnetic nanofibers with beads, tailored by a reduction of the polymer content in the spinning solution, as evaluated in detail in a former study [59].…”

Section: Introductionmentioning

confidence: 99%

Magnetic Properties of Electrospun Magnetic Nanofiber Mats after Stabilization and Carbonization

et al. 2020

View full text Add to dashboard Cite

Magnetic nanofibers are of great interest in basic research, as well as for possible applications in spintronics and neuromorphic computing. Here we report on the preparation of magnetic nanofiber mats by electrospinning polyacrylonitrile (PAN)/nanoparticle solutions, creating a network of arbitrarily oriented nanofibers with a high aspect ratio. Since PAN is a typical precursor for carbon, the magnetic nanofiber mats were stabilized and carbonized after electrospinning. The magnetic properties of nanofiber mats containing magnetite or nickel ferrite nanoparticles were found to depend on the nanoparticle diameters and the potential after-treatment, as compared with raw nanofiber mats. Micromagnetic simulations underlined the different properties of both magnetic materials. Atomic force microscopy and scanning electron microscopy images revealed nearly unchanged morphologies after stabilization without mechanical fixation, which is in strong contrast to pure PAN nanofiber mats. While carbonization at 500 °C left the morphology unaltered, as compared with the stabilized samples, stronger connections between adjacent fibers were formed during carbonization at 800 °C, which may be supportive of magnetic data transmission.

show abstract

“…28 Very recently, Wu's group demonstrated reliable and uniform analogue switching behaviors in the 2048 1T1M crossbar array of TiN/TaO x /HfO x /TiN devices. 29 By integrating eight 2048-cell 1T1M crossbar arrays on a printed circuit board (PCB) and a eld-programmable gate array evaluation board (ZC706, Xilinx), they successfully built a ve-layer memristorbased CNN and performed MNIST image recognition with a high accuracy of more than 96 per cent. They demonstrated that the energy efficiency in memristor-based CNN neuromorphic systems is two orders of magnitude greater than that of the state-of-the-art graphics-processing units.…”

Section: T1mmentioning

confidence: 99%

Research progress on solutions to the sneak path issue in memristor crossbar arrays

et al. 2020

View full text Add to dashboard Cite

show abstract

Fully hardware-implemented memristor convolutional neural network

Cited by 1,560 publications

References 33 publications

Implementation of Dropout Neuronal Units Based on Stochastic Memristive Devices in Neural Networks with High Classification Accuracy

Implementation of Dropout Neuronal Units Based on Stochastic Memristive Devices in Neural Networks with High Classification Accuracy

Magnetic Properties of Electrospun Magnetic Nanofiber Mats after Stabilization and Carbonization

Research progress on solutions to the sneak path issue in memristor crossbar arrays

Contact Info

Product

Resources

About