Energy-Efficient Implementation of Generative Adversarial Networks on
Passive RRAM Crossbar Arrays
Abstract
Generative algorithms such as GANs are at the cusp of next revolution in
the field of unsupervised learning and large-scale artificial data
generation. However, the adversarial (competitive) co-training of the
discriminative and generative networks in GAN makes them computationally
intensive and hinders their deployment on the resource-constrained IoT
edge devices. Moreover, the frequent data transfer between the
discriminative and generative networks during training significantly
degrades the efficacy of the von-Neumann GAN accelerators such as those
based on GPU and FPGA. Therefore, there is an urgent need for
development of ultra-compact and energy-efficient hardware accelerators
for GANs. To this end, in this work, we propose to exploit the passive
RRAM crossbar arrays for performing key operations of a fully-connected
GAN: (a) true random noise generation for the generator network, (b)
vector-by-matrix-multiplication with unprecedented energy-efficiency
during the forward pass and backward propagation and (C) in-situ
adversarial training using a hardware friendly Manhattan’s rule. Our
extensive analysis utilizing an experimentally calibrated phenomological
model for passive RRAM crossbar array reveals an unforeseen trade-off
between the accuracy and the energy dissipated while training the GAN
network with different noise inputs to the generator. Furthermore, our
results indicate that the spatial and temporal variations and true
random noise, which are otherwise undesirable for memory application,
boost the energy-efficiency of the GAN implementation on passive RRAM
crossbar arrays without degrading its accuracy.