Abstract
Generative Adversarial Network (GAN) requires extensive computing resources making its implementation in edge devices with conventional microprocessor hardware a slow and difficult, if not impossible task. In this paper, we propose to accelerate these intensive neural computations using memristive neural networks in analog domain. The implementation of Analog Memristive Deep Convolutional GAN (AM-DCGAN) using Generator as deconvolutional and Discriminator as convolutional memristive neural network is presented. The system is simulated at circuit level with 1.7 million memristor devices taking into account memristor non-idealities, device and circuit parameters. The design is modular with crossbar arrays having a minimum average power consumption per neural computation of 47nW. The design exclusively uses the principles of neural network dropouts resulting in regularization and lowering the power consumption. The SPICE level simulation of GAN is performed with 0.18 μm CMOS technology and WOx memristive devices with RON = 40 kΩ and ROFF = 250 kΩ, threshold voltage 0.8 V and write voltage at 1.0 V.
Highlights
Generative Adversarial Network (GAN) requires extensive computing resources making its implementation in edge devices with conventional microprocessor hardware a slow and difficult, if not impossible task
This paper proposes the first implementation of fully analog memristive hardware implementation of AM-Deep Convolutional GAN (DCGAN)
Up to 10% of noise in the input images, that can be the result of the interface between image sensor and the proposed circuit in near-sensor processing, can be tolerated
Summary
Generative Adversarial Network (GAN) requires extensive computing resources making its implementation in edge devices with conventional microprocessor hardware a slow and difficult, if not impossible task. The main contributions of the paper are the following: (1) the design and circuit level implementation of a Deep Convolutional Generative Adversarial Network with memristive crossbar arrays having 1.7 million memristor devices, (2) a scalable design that provides high robustness to conductance variability, circuit parasitics and device non-idealities, which is verified in the paper, and (3) as the presented architecture is analog, it can be integrated into the chip for near-sensor processing without the analog to digital conversion for both training and image generation This can increase the processing and training speed and reduce the power consumption, comparing to traditional CPU- and GPU-based GAN processing[20], making the architecture applicable for edge processing
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have