The Research On Resistive Random-access Memory Array Based Neural Network

The proposed generative adversarial network meets the need of research and application in many fields,but it still has some problems.Such as instability or even difficult to converge,and pattern collapse.In order to solve these problems,the sample can be noise-added before being sent to the discriminator when training the GAN.In the naive generative adversarial network,noise is added to each sample one by one during training.This operation consumes resources and the added noise is pseudorandom noise.Memristor devices can act as synaptic weights in brain-inspired computing,with low power consumption and high speed,but their computational accuracy is lower than that of floating-point-based CPU/GPU due to their inherent nonideal factors.In this paper,the memristor generative adversarial network is realized by the memristor array.In the training,the intrinsic random noise of the analog RRAM device is used as the input of the neural network,which improves the diversity of the generated patterns and solves the aforementioned problems faced by generative adversarial network.The specific work of this paper includes:1.The intrinsic random noise,the unique behavior of read and write noise in analog RRAM devices,is studied.And a memristor model for simulating neural networks is established.Through the simulation of memristive GAN,the influence of read and write noise on the performance of memristive GAN is analyzed,and the impact mode of read and write noise in neural network is also analyzed.2.Aiming at the characteristics of RRAM,an improved training method of memristive GAN is proposed to mitigate the influence of excessive noise of RRAM.At the same time,conductance update method with verification is used,which tolerates the noise of the device,and accelerates the convergence speed of the memristive GAN.3.For the first time,the experiment on a 1 Kb analog RRAM array verify that a solution to implement memristive GAN using RRAM devices is feasible.After online training,memristive GAN can generate different digital patterns,solving the problem of the loss of sample diversity and the vanishing of gradient.The work of this paper proves that RRAM is suitable for the applications of GAN,and it also paves a new way to take advantage of the non-ideal effects of RRAM devices.