This research enhances GANs by introducing a novel Parametric Mish (PMish) activation function, MMD-GAN repulsive loss incorporated neural architecture search strategy, and an adaptive rank decomposition method for network compression. The study improves GAN performance and stability with Modified MMD-GAN repulsive loss, demonstrating significant results on datasets like CIFAR-10, CIFAR-100, STL-10, and CelebA. Key contributions include the introduction of the PMish activation function, modified MMD-GAN rep loss, training strategy, and efficient model compression using ARD for deployment on resource-constrained devices.

MMD-AdversarialNAS is a framework that utilizes Neural Architecture Search (NAS) and Maximum Mean Discrepancy (MMD) repulsive loss to optimize GAN architectures for image generation while incorporating tensor decomposition to significantly reduce model complexity and storage footprint.

ProGAN (Progressive Growing of GANs) is a method for training GANs that progressively increases the resolution of generated images, starting from low resolution and growing layers incrementally for better stability and image quality.

Human Action Recognition seeks to automatically identify human actions from a series of observations (image frames), typically a video. The advent of Deep Neural Networks has significantly advanced the state-of-the-art in this field. Our research aims to enhance action recognition performance by leveraging data augmentation, specifically by generating synthetic videos using GANs.