Deep neural network representation and Generative Adversarial Learning

Generative Adversarial Networks (GANs) have proven to be efficient systems for data generation and other machine learning tasks. They owe their success to a minimax learning concept initially proposed by Schmidhuber (1990) to implement Artificial Curiosity. Two learning networks, a generator and an evaluator or discriminator, compete with each other in a zero-sum game. Despite their obvious advantages and their application to a wide range of domains, GANs have yet to overcome several challenges such as non-convergence, overfitting, mode collapse, amongst others. New advancements in deep representation learning (RL) can help improve the learning process in GenerativeAdversarial Learning (GAL). For instance, RL can help address issues such as dataset bias and identify a set of features that are well suited for a given task.