Publication

Abstract : In this chapter, we aim to scale-up the operation of GAN using recurrent neural network (RNN). The GAN is designed with two deep neural network models, where the first network is called a generator that generates perceptually synthetic samples drawn through real data distribution. The first network recursively learns the network and transforms the noise vectors into samples. The second network is called a discriminator that accesses both real and synthetic samples and provides classification between these two data. The discriminator uses the same RNN to classify the generated and real data samples. The aim of using RNN in the GAN structure is to delimit the error rate using its time series prediction based on its past inputs. It further uses its neighborhood relationship between the samples to generate the target output and error generated. The errors are then propagated in the backward direction over the GAN to update the network weights to estimate the output. The RNN generator, on the other hand, reduces the probability of the RNN discriminator in identifying the false generated samples and increasing the probability of discriminator in identifying correctly the real samples. The objective function in RNN is designed in such a way that its gradient operator for the false samples is quite far from the decision boundary of RNN discriminator, thereby producing the increasing true classification rate. The experiments carried out on the real-world time series dataset show the results of accurate classification (97.01%) with increased false detection rate than benchmark GAN method.