In recent times, the Generative Adversarial Network (GAN) has gained great attention in academia and industry. Also, it ascertains immense useful applicability across multiple domains. GANs are engrossing developments in artificial intelligence, most importantly, the unsupervised learning paradigm. The main significant characteristics of GAN are learning high-dimensional, complex real data distribution and excellent data generation capability. Abstractly, GAN implements unsupervised learning by enchanting a supervised learning approach and generating synthetic data.

   GANs are inspired by the game theory; the generator and discriminator compete with each other to learn deep representations without extra training data for the distribution of stunningly realistic samples. The image and vision field admits more popularity of GANs researchers as it has outstanding advantages over the other generative models.

Diverse Variants of GAN:

   Owing to the inadequateness of original GANs, several variants of GANs models were conceptualized and developed. Most adapted and recent variants of GAN are discussed below;

   Conditional Generative Adversarial Networks (CGAN) - CGAN is the conditional version of GAN, constructed by merely provisioning the extra auxiliary information into the form CGAN. CGAN controls the generation of the image with its conditional variables. CGAN models edict the type of data generated via the applied condition and produces a general framework for different applications.

   Deep Convolutional Generative Adversarial Networks (DCGAN) - DCGAN is a new class of convolutional neural networks (CNN) that significantly demonstrates a steady GAN training procedure and achieves high performance in superior quality sharp images generation and high-resolution image generation tasks.

   Bidirectional GAN - BiGANs are closely similar to autoencoders with a loss function and provide rich representations in unsupervised learning. BiGANs are suspicious of the domain of the data and efforts to learn an inverse mapping by protruding the data back into the latent space.

   Super-Resolution Generative Adversarial Network (SRGAN) - SRGAN utilizes a very deep convolutional network through residual blocks with adversarial loss and a feature loss. Experimentally a high performance on public datasets is manifested with the help of SRGAN.

   Information Maximizing Generative Adversarial Networks (InfoGAN) - InfoGAN is an information-theoretic extension of GAN and learns the disentangled representations in an entirely supervised manner. The uniqueness of the InfoGAN is the instigation of regularization for capturing shared information among interpretable variables. InfoGANs are preferred non-complex datasets.

   Wasserstein GAN (WGAN) -WGAN is the most straightforward category of GAN and uses an alternate cost function to stabilize the GAN training and imparts better outcomes by resolving the training problems such as vanishing gradient and the mode collapse problems very well.

   BigGAN - Big GAN is one of the recent best models due to its outstanding, large scale, indistinguishable, and high-fidelity diverse image generation capacity. BigGAN provides extremely detailed images with remarkable performance in training bigger neural networks even with more parameters.

Notable Applications of GAN:

   GAN explores unbound and wide applications in various areas for their effective use and is relatively a very new technology, lots of research is still being accomplished to enrich and create better and preponderant GAN-based models. GANs are an eminently amazing generative model in the synthesis of realistic-looking samples. Thence, the advantages conduct GAN to be employed in different fields of Computer Vision (CV), Natural Language Processing (NLP), and Artificial Intelligence (AI). Here, GAN applications in several domains, such as image, audio, and video, are highlighted;

Image Applications:

   • GAN is effectively used for automatic character generation applications, which include font, and handwritten text, and some of the recent models are Conditional Least Square GAN (LSCGAN), Multi-Scale Multi-Class Conditional GAN (MCMS-CGAN), Handwritten GAN (HW-GAN), a Semi-Supervised GAN (SS-GAN).

   • Cost-efficient anime-character generation is realizable using GAN with less amount of artistic skills. Several efforts have developed, such as Chainer-DCGAN, Illustration-Style Reproduction DCGAN Deep Regret Analytic GAN (DRAGAN) Progressive Structure-Conditional GAN (PS-CGAN).

   • By applying GAN, a dominant performance is realized in several computer vision-based blending of images together in visual communication or automatic photo editing tasks. Gaussian Poisson GAN (GPGAN) and Color Consistent GAN (GCC-GAN) are widely used proposed image blending methods for high-resolution images.

   • Geometrically, Progressive-Growing Generative Adversarial Networks (PGGAN) and Exemplar GAN (Ex-GAN) are utilized for image inpainting tasks by aggregating local and global information.

   • Photo-editing, computer-aided design, and image synthesis for generating high-quality photo-realistic images from the text is tremendous application. Attentional GAN (AttnGAN), Stacked GAN (StackGAN), and Auxiliary Classifier GAN (AC-GAN) are lately applied for such applications.

   • In human pose synthesis, Deformable GAN generates a person-s image based on appearance and poses information, and eliminates distortion between the synthesized and the ground-truth images.

   • Research has been made on visual saliency detection using Saliency Prediction GAN (Sal-GAN) De-noising Saliency Prediction GAN (DSAL-GAN) is beneficial to produce accurate outcomes.

   • For 3D image synthesis, GANs are applied to generate 3D objects via advances in 3-dimensional (3D) volumetric convolution networks. 3D-GAN broadens the system from 2D-GAN to 3D-GAN. Projective GAN (Pr-GAN) is proposed to gather 3D objects from multiple 2D views of an object.

   • In medical image applications, GAN is a well-accepted contemporary area of research, and many works have evolved Segmentation GAN (SegAN), Medical GAN (Med-GAN), and online doctor recommendation framework (DR-GAN)

   • Texture and sketch synthesis image to image translation is another recent image-based applicative task of GAN.

Audio Applications:

   • GAN has progressive success in the synthesis of data in music generation, dialogue systems, and machine translation

   • Variational Auto-encoding WGAN (VAE-WGAN) is developed for the voice conversion system.

   • For music generation, GAN models such as Continuous Recurrent Neural Network GAN (CRNN-GAN), Sequence GAN (Seq-GAN), and Object Reinforced GAN (OR-GAN) are established.

   • A novel high-quality language generation is achieved by the recent GAN model, Ranker GAN (Rank-GAN)

Video Applications:

   • Prophesy the object motions is the main issue of video generation tasks in computer vision. GAN-based video generation (VGAN) and Motion and Content GAN (MoCo-GAN) are developed to solve such issues in video generation.

   • Disentangled Representation Net (DRNET) is evolved for image representations from the video frames

   • Generation of high-resolution videos with high reliability is accomplished by Dual Video Discriminator GAN (DVDGAN)

   • Future video prediction is performed using Pose-GAN (PGAN) and Dual-Motion GAN (DMGAM)

Multifarious Applications:

   GANs have also been utilized in many other implementation areas such as malware detection - chess game playing - network pruning - spatial representation learning - mobile user profiling - data augmentation - heterogeneous information networks - privacy-preserving - social robot - cipher cracking - auxiliary automatic driving - continual learning - molecule development in oncology - GANs for finance - GANs for textile - GANs for e-commerce - GANs for fluid-flow.

Limitations of GAN:

   Nevertheless, GANs are state of the art model that remains as highly inconstant due to the following causes.,

   • Evaluation metrics - Unavailability of a standard defined evaluation metric because of this consideration, evaluation of GAN models is an active research area.

   • Mode collapse - Mode collapse is another challenge in GAN. Owing to the difficulty, the GAN model fails to progress optimization, produce identical output, and leads to a low diversity

   • Nash equilibrium - This challenge yields unstable GAN. As interdependence of discriminator and generator, need to achieve Nash equilibrium which is very complex for both generator and discriminator.

   • Vanishing Gradient - Vanishing gradient appears in generator with small gradients and yields in discriminating the generated data as fake with a better probability but further leads to instability.

   • Internal covariate shift -Internal Covariate Shift occurs during the input distribution of network activation varies due to updating parameters in preceding layers, further leading to high computation time and training cost.

Future Research Areas of GAN:

   In perceptive of the future, the impediments of GANs will be resolved innovatively, and a new approach will be evolved. By ameliorating the architecture of the network and algorithms of GAN, a more powerful generative model to generate images, audios, videos, and texts that are complicated for humans to distinguish. Some of the future developmental trends that need to contend under the area of GAN are:

   • One of the main future directions in GAN is to merge GAN and semi-supervised learning.

   • Some other combined concepts for future scopes are embedded GANs into imitation learning, inter-fused GANs with policy gradient, and describing the relationship between GANs and actor-critic algorithms.

   • Developing GANs for more robustness in adversarial attacks is another research scope.

   • New families of Integral Probability Metrics (IPMs) for training GANs must focus on enhancement as new divergences.

   • GANs to generate data with the measure of uncertainty of the well-trained generator is another captivating future issue.

   • Unlocking the potentiality of GANs for NLP and hashing in terms of discrete data.