Troika GAN vs Decoupled GAN: An Investigation on the Impact of Subnetwork Weight Sharing for Data Augmentation

Notable advancements in the field of computer vision have transpired through the application of Generative Adversarial Networks (GANs). A new GAN variant, the Troika GAN (T-GAN), was recently proposed for data augmentation and was shown to be superior to the Coupled GAN (CoGAN) and the classic techniques of rotation and affine transformation. This paper describes our further investigation on T-GAN, specifically the impact of its subnetworks weight sharing. We decoupled the weight-sharing subnetworks of T-GAN to form three independent GANs, which we refer to collectively as the Decoupled GAN and whose weights are trained separately. We then used T-GAN and the Decoupled GAN to augment a set of words with limited instances from the IAM Handwriting Database. The resulting augmented datasets were applied to train the three types of Artificial Neural Network (ANN) classifiers: Vanilla ANN, Deep ANN, and Convolutional Neural Network (CNN). Results showed that the best accuracies from each of the 3 classifier types were obtained when these were trained with datasets augmented by a T-GAN. For example, the CNN classifier registered 89.76% as its best performance using T-GAN while recording only 87.47% accuracy from utilizing Decoupled GAN. A paired t-test between the 10-fold cross-validation results of these yielded a statistically significant p-value of 0.0075 in favor of the T-GAN augmentation. This clearly indicates that the sharing of weights is a vital factor in the generation of better synthetic data. With its significant impact on improving handwriting classification networks, T-GAN can be an ideal data augmentation approach to build robust systems where there is a scarcity of training dataset instances.