Authors - Ankit Shah, Hardik M. Patel Abstract - Generative Adversarial Networks (GANs) have revolutionized data augmentation by generating realistic and diverse synthetic data, significantly enhancing the performance of machine learning models. This review evaluates the efficacy of GAN-based augmentation compared to traditional methods across various datasets, including MNIST, CIFAR-10, and diabetic retinopathy images. Using architectures such as DCGAN, WGAN-GP, and StyleGAN, our experiments showed substantial performance improvements: CNN accuracy on CIFAR-10 increased from 82.0% to 87.5%, and ResNet-50 accuracy on diabetic retinopathy images rose from 75.0% to 87.0%. Statistical analyses confirmed the significance of these gains. Despite challenges like computational costs and training instability, GAN-based augmentation proves superior in addressing data scarcity and enhancing model robustness. Future research should focus on optimizing GAN training, integrating hybrid models, and exploring ethical considerations. The results underscore GANs' potential in advancing machine learning applications, particularly in complex and data-scarce domains.
Open Zoom
